Professional-Machine-Learning-Engineer Questions and Answers

The best option for implementing the processing steps so that they run at serving time, minimizing code changes and infrastructure maintenance, and deploying the model into production as quickly as possible, is to use the Predictor interface to implement a custom prediction routine. Build the custom container, upload the container to Vertex AI Model Registry, and deploy it to a Vertex AI endpoint. This option allows you to leverage the power and simplicity of Vertex AI to serve your XGBoost model with minimal effort and customization. Vertex AI is a unified platform for building and deploying machine learning solutions on Google Cloud. Vertex AI can deploy a trained XGBoost model to an online prediction endpoint, which can provide low-latency predictions for individual instances. A custom prediction routine (CPR) is a Python script that defines the logic for preprocessing the input data, running the prediction, and postprocessing the output data. A CPR can help you customize the prediction behavior of your model, and handle complex or non-standard data formats. A CPR can also help you minimize the code changes, as you only need to write a few functions to implement the prediction logic. A Predictor interface is a class that inherits from the base class aiplatform.Predictor, and implements the abstract methods predict() and preprocess(). A Predictor interface can help you create a CPR by defining the preprocessing and prediction logic for your model. A container image is a package that contains the model, the CPR, and the dependencies. A container image can help you standardize and simplify the deployment process, as you only need to upload the container image to Vertex AI Model Registry, and deploy it to Vertex AI Endpoints. By using the Predictor interface to implement a CPR, building the custom container, uploading the container to Vertex AI Model Registry, and deploying it to a Vertex AI endpoint, you can implement the processing steps so that they run at serving time, minimize code changes and infrastructure maintenance, and deploy the model into production as quickly as possible1.

The other options are not as good as option C, for the following reasons:

• Option A: Using FastAPI to implement an HTTP server, creating a Docker image that runs your HTTP server, and deploying it on your organization’s GKE cluster would require more skills and steps than using the Predictor interface to implement a CPR, building the custom container, uploading the container to Vertex AI Model Registry, and deploying it to a Vertex AI endpoint. FastAPI is a framework for building web applications and APIs in Python. FastAPI can help you implement an HTTP server that can handle prediction requests and responses, and perform data preprocessing and postprocessing. A Docker image is a package that contains the model, the HTTP server, and the dependencies. A Docker image can help you standardize and simplify the deployment process, as you only need to build and run the Docker image. GKE is a service that can create and manage Kubernetes clusters on Google Cloud. GKE can help you deploy and scale your Docker image on Google Cloud, and provide high availability and performance. However, using FastAPI to implement an HTTP server, creating a Docker image that runs your HTTP server, and deploying it on your organization’s GKE cluster would require more skills and steps than using the Predictor interface to implement a CPR, building the custom container, uploading the container to Vertex AI Model Registry, and deploying it to a Vertex AI endpoint. You would need to write code, create and configure the HTTP server, build and test the Docker image, create and manage the GKE cluster, and deploy and monitor the Docker image. Moreover, this option would not leverage the power and simplicity of Vertex AI, which can provide online prediction natively integrated with Google Cloud services2.
• Option B: Using FastAPI to implement an HTTP server, creating a Docker image that runs your HTTP server, uploading the image to Vertex AI Model Registry, and deploying it to a Vertex AI endpoint would require more skills and steps than using the Predictor interface to implement a CPR, building the custom container, uploading the container to Vertex AI Model Registry, and deploying it to a Vertex AI endpoint. FastAPI is a framework for building web applications and APIs in Python. FastAPI can help you implement an HTTP server that can handle prediction requests and responses, and perform data preprocessing and postprocessing. A Docker image is a package that contains the model, the HTTP server, and the dependencies. A Docker image can help you standardize and simplify the deployment process, as you only need to build and run the Docker image. Vertex AI Model Registry is a service that can store and manage your machine learning models on Google Cloud. Vertex AI Model Registry can help you upload and organize your Docker image, and track the model versions and metadata. Vertex AI Endpoints is a service that can provide online prediction for your machine learning models on Google Cloud. Vertex AI Endpoints can help you deploy your Docker image to an online prediction endpoint, which can provide low-latency predictions for individual instances. However, using FastAPI to implement an HTTP server, creating a Docker image that runs your HTTP server, uploading the image to Vertex AI Model Registry, and deploying it to a Vertex AI endpoint would require more skills and steps than using the Predictor interface to implement a CPR, building the custom container, uploading the container to Vertex AI Model Registry, and deploying it to a Vertex AI endpoint. You would need to write code, create and configure the HTTP server, build and test the Docker image, upload the Docker image to Vertex AI Model Registry, and deploy the Docker image to Vertex AI Endpoints. Moreover, this option would not leverage the power and simplicity of Vertex AI, which can provide online prediction natively integrated with Google Cloud services2.
• Option D: Using the XGBoost prebuilt serving container when importing the trained model into Vertex AI, deploying the model to a Vertex AI endpoint, working with the backend engineers to implement the pre- and postprocessing steps in the Golang backend service would not allow you to implement the processing steps so that they run at serving time, and could increase the code changes and infrastructure maintenance. A XGBoost prebuilt serving container is a container image that is provided by Google Cloud, and contains the XGBoost framework and the dependencies. A XGBoost prebuilt serving container can help you deploy a XGBoost model without writing any code, but it also limits your customization options. A XGBoost prebuilt serving container can only handle standard data formats, such as JSON or CSV, and cannot perform any preprocessing or postprocessing on the input or output data. If your input data requires any transformation or normalization before running the prediction, you cannot use a XGBoost prebuilt serving container. A Golang backend service is a service that is implemented in Golang, a programming language that can be used for web development and system programming. A Golang backend service can help you handle the prediction requests and responses from the frontend, and communicate with the Vertex AI endpoint. However, using the XGBoost prebuilt serving container when importing the trained model into Vertex AI, deploying the model to a Vertex AI endpoint, working with the backend engineers to implement the pre- and postprocessing steps in the Golang backend service would not allow you to implement the processing steps so that they run at serving time, and could increase the code changes and infrastructure maintenance. You would need to write code, import the trained model into Vertex AI, deploy the model to a Vertex AI endpoint, implement the pre- and postprocessing steps in the Golang backend service, and test and monitor the Golang backend service. Moreover, this option would not leverage the power and simplicity of Vertex AI, which can provide online prediction natively integrated with Google Cloud services2.

References:

• Preparing for Google Cloud Certification: Machine Learning Engineer, Course 3: Production ML Systems, Week 2: Serving ML Predictions
• Google Cloud Professional Machine Learning Engineer Exam Guide, Section 3: Scaling ML models in production, 3.1 Deploying ML models to production
• Official Google Cloud Certified Professional Machine Learning Engineer Study Guide, Chapter 6: Production ML Systems, Section 6.2: Serving ML Predictions
• Custom prediction routines
• Using pre-built containers for prediction
• Using custom containers for prediction

• Option A is not the best answer because it requires storing the pickled model in Cloud Storage, which may incur additional cost and latency for loading the model. It also requires building a Flask-based app, which may not be necessary for a simple data preprocessing step.
• Option B is not the best answer because it requires building a Flask-based app, which may not be necessary for a simple data preprocessing step. It also requires packaging the app and the pickled model in a custom container image, which may increase the size and complexity of the image.
• Option C is not the best answer because it requires packaging the pickled model in a custom container image, which may increase the size and complexity of the image. It also does not leverage the Vertex built-in container image, which may provide some optimizations and integrations for XGBoost models.
• Option D is the best answer because it leverages the Vertex built-in container image, which may provide some optimizations and integrations for XGBoost models. It also allows storing the pickled model in Cloud Storage, which may reduce the size and complexity of the image. It also allows building a custom predictor class based on XGBoost Predictor from the Vertex AI SDK, which may simplify the data preprocessing step and the prediction logic.

AutoML Vision Edge is a service that allows you to create custom image classification and object detection models that can run on edge devices, such as mobile phones, tablets, or IoT devices1. AutoML Vision Edge offers four types of models that vary in size, accuracy, and latency: mobile-versatile-1, mobile-low-latency-1, mobile-high-accuracy-1, and mobile-core-ml-low-latency-12. Each model has its own trade-offs and use cases, depending on the device specifications and the application requirements.

For the use case of building an ML model to reduce the amount of time spent by quality control inspectors checking for product defects, the best model to use is the AutoML Vision Edge mobile-low-latency-1 model. This model is optimized for fast inference on mobile devices, with a latency of less than 50 milliseconds on a Pixel 1 phone2. Faster defect detection is a priority for the toy manufacturer, and the factory does not have reliable Wi-Fi, so a low-latency model that can run on the device without internet connection is ideal. The mobile-low-latency-1 model also has a small size of less than 4 MB, which makes it easy to deploy and update2. The mobile-low-latency-1 model has a slightly lower accuracy than the mobile-high-accuracy-1 model, but it is still suitable for most image classification tasks2. Therefore, the AutoML Vision Edge mobile-low-latency-1 model is the best option for this use case.

References:

• AutoML Vision Edge documentation
• AutoML Vision Edge model types

The best business metric to track to measure the model’s performance is user engagement as measured by the number of battles played daily per user. This metric reflects the main goal of the model, which is to enhance the user experience and satisfaction by creating balanced and fair battles. If the model is successful, it should increase the user retention and loyalty, as well as the word-of-mouth and referrals. This metric is also easy to measure and interpret, as it can be directly obtained from the user activity data.

The other options are not optimal for the following reasons:

• A. Average time players wait before being assigned to a team is not a good metric, as it does not capture the quality or outcome of the battles. It only measures the efficiency of the model, which is not the primary objective. Moreover, this metric can be influenced by external factors, such as the availability and demand of players, the network latency, and the server capacity.
• B. Precision and recall of assigning players to teams based on their predicted versus actual ability is not a good metric, as it is difficult to measure and interpret. It requires having a reliable and consistent way of estimating the player’s ability, which can be subjective and dynamic. It also requires having a ground truth label for each assignment, which can be costly and impractical to obtain. Moreover, this metric does not reflect the user feedback or satisfaction, which is the ultimate goal of the model.
• D. Rate of return as measured by additional revenue generated minus the cost of developing a new model is not a good metric, as it is not directly related to the model’s performance. It measures the profitability of the model, which is a secondary objective. Moreover, this metric can be affected by many other factors, such as the market conditions, the pricing strategy, the marketing campaigns, and the competition.

References:

• Professional ML Engineer Exam Guide
• Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate
• Google Cloud launches machine learning engineer certification
• How to measure user engagement
• How to choose the right metrics for your machine learning model

• Option A is incorrect because creating a collaborative filtering system that recommends articles to a user based on the user’s past behavior is not the best approach to suggest articles that are similar to the articles they are currently reading. Collaborative filtering is a method of recommendation that uses the ratings or preferences of other users to predict the preferences of a target user1. However, this method does not consider the content or features of the articles, and may not be able to find articles that are similar in terms of topic, style, or sentiment.
• Option B is correct because encoding all articles into vectors using word2vec, and building a model that returns articles based on vector similarity is a suitable approach to suggest articles that are similar to the articles they are currently reading. Word2vec is a technique that learns low-dimensional and dense representations of words from a large corpus of text, such that words that are semantically similar have similar vectors2. By applying word2vec to the articles, we can obtain vector representations of the articles that capture their meaning and usage. Then, we can use a similarity measure, such as cosine similarity, to find articles that have similar vectors to the current article3.
• Option C is incorrect because building a logistic regression model for each user that predicts whether an article should be recommended to a user is not a feasible approach to suggest articles that are similar to the articles they are currently reading. Logistic regression is a supervised learning method that models the probability of a binary outcome (such as recommend or not) based on some input features (such as user profile or article content)4. However, this method requires a large amount of labeled data for each user, which may not be available or scalable. Moreover, this method does not directly measure the similarity between articles, but rather the likelihood of a user’s preference.
• Option D is incorrect because manually labeling a few hundred articles, and then training an SVM classifier based on the manually classified articles that categorizes additional articles into their respective categories is not an effective approach to suggest articles that are similar to the articles they are currently reading. SVM (support vector machine) is a supervised learning method that finds a hyperplane that separates the data into different classes (such as news categories) with the maximum margin5. However, this method also requires a large amount of labeled data, which may be costly and time-consuming to obtain. Moreover, this method does not account for the fine-grained similarity between articles within the same category, or the cross-category similarity between articles from different categories.

References:

• Collaborative filtering
• Word2vec
• Cosine similarity
• Logistic regression
• SVM

According to the official exam guide1, one of the skills assessed in the exam is to “design, build, and productionalize ML models to solve business challenges using Google Cloud technologies”. Vertex AI AutoML Forecasting2 is a service that allows you to train and deploy custom time-series forecasting models for batch prediction. Vertex AI AutoML Forecasting simplifies the model development process by providing a graphical user interface and a no-code approach. You can use Vertex AI AutoML Forecasting to train a model by using your tabular data, and specify the target variable, the covariates, and the time variable. Vertex AI AutoML Forecasting automatically handles the feature engineering, model selection, and hyperparameter tuning. Therefore, option D is the best way to maximize the speed of model development and testing for the given use case. The other options are not relevant or optimal for this scenario. References:

• Professional ML Engineer Exam Guide
• Vertex AI AutoML Forecasting
• Google Professional Machine Learning Certification Exam 2023
• Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

• Option A is not the best answer because it requires modifying the pipeline to use the Vertex AI Feature Store, which may not be feasible or necessary for recovering the data that the model was trained on. The Vertex AI Feature Store is a service that helps you manage, store, and serve feature values for your machine learning models1, but it is not designed for storing the raw data or the TFRecord files.
• Option B is the best answer because it leverages the lineage feature of Vertex AI Metadata, which is a service that helps you track and manage the metadata of your machine learning workflows, such as datasets, models, metrics, and parameters2. The lineage feature allows you to view the relationships and dependencies among the artifacts and executions in your pipeline, and trace back the origin and history of any artifact3. By using the lineage feature, you can find the model artifact, determine the version of the model, identify the step that creates the data copy, and search in the metadata for its location.
• Option C is not the best answer because it relies on the logging features in the Vertex AI endpoint, which may not be accurate or reliable for finding the data copy. The logging features in the Vertex AI endpoint help you monitor and troubleshoot the online predictions made by your deployed models, but they do not provide information about the training data or the pipeline steps4. Moreover, the timestamp of the model deployment may not match the timestamp of the pipeline run, as there may be delays or errors in the deployment process.
• Option D is not the best answer because it requires finding the job ID in Vertex AI Training, which may not be easy or straightforward. Vertex AI Training is a service that helps you train your custom models on Google Cloud, but it does not provide a direct way to link the training job to the model version or the pipeline run. Moreover, searching in the logs of the job may not reveal the location of the data copy, as the logs may only contain information about the training process and the metrics.

References:

• 1: Introduction to Vertex AI Feature Store | Vertex AI | Google Cloud
• 2: Introduction to Vertex AI Metadata | Vertex AI | Google Cloud
• 3: View lineage for ML workflows | Vertex AI | Google Cloud
• 4: Monitor online predictions | Vertex AI | Google Cloud
• [5]: Train custom models | Vertex AI | Google Cloud

An input pipeline is a way to prepare and feed data to a machine learning model for training or inference. An input pipeline typically consists of several steps, such as reading, parsing, transforming, batching, and prefetching the data. An input pipeline can improve the performance and efficiency of the model, as it can handle large and complex datasets, optimize the data processing, and reduce the latency and memory usage1.

For the use case of developing an input pipeline for an ML training model that processes images from disparate sources at a low latency, the best option is to convert the images into TFRecords, store the images in Cloud Storage, and then use the tf.data API to read the images for training. This option involves using the following components and techniques:

• TFRecords: TFRecords is a binary file format that can store a sequence of data records, such as images, text, or audio. TFRecords can help to compress, serialize, and store the data efficiently, and reduce the data loading and parsing time. TFRecords can also support data sharding and interleaving, which can improve the data throughput and parallelism2.
• Cloud Storage: Cloud Storage is a service that allows you to store and access data on Google Cloud. Cloud Storage can help to store and manage large and distributed datasets, such as images from different sources, and provide high availability, durability, and scalability. Cloud Storage can also integrate with other Google Cloud services, such as Compute Engine, AI Platform, and Dataflow3.
• tf.data API: tf.data API is a set of tools and methods that allow you to create and manipulate data pipelines in TensorFlow. tf.data API can help to read, transform, batch, and prefetch the data efficiently, and optimize the data processing for performance and memory. tf.data API can also support various data sources and formats, such as TFRecords, CSV, JSON, and images.

By using these components and techniques, the input pipeline can process large datasets of images from disparate sources that do not fit in memory, and provide low latency and high performance for the ML training model. Therefore, converting the images into TFRecords, storing the images in Cloud Storage, and using the tf.data API to read the images for training is the best option for this use case.

References:

• Build TensorFlow input pipelines | TensorFlow Core
• TFRecord and tf.Example | TensorFlow Core
• Cloud Storage documentation | Google Cloud
• [tf.data: Build TensorFlow input pipelines | TensorFlow Core]

TFRecord is a binary file format that stores your data as a sequence of binary strings1. TFRecord files are efficient, scalable, and easy to process1. Sharding is a technique that splits a large file into smaller files, which can improve parallelism and performance2. Dataflow is a service that allows you to create and run data processing pipelines on Google Cloud3. Dataflow can create sharded TFRecord files from your images in a Cloud Storage directory4.

tf.data.TFRecordDataset is a class that allows you to read and parse TFRecord files in TensorFlow. You can use this class to create a tf.data.Dataset object that represents your input data for training. tf.data.Dataset is a high-level API that provides various methods to transform, batch, shuffle, and prefetch your data.

Vertex AI Training is a service that allows you to train your custom models on Google Cloud using various hardware accelerators, such as GPUs. Vertex AI Training supports TensorFlow models and can read data from Cloud Storage. You can use Vertex AI Training to train your image classification model by using a V100 GPU, which is a powerful and fast GPU for deep learning.

References:

• TFRecord and tf.Example | TensorFlow Core
• Sharding | TensorFlow Core
• Dataflow | Google Cloud
• Creating sharded TFRecord files | Google Cloud
• [tf.data.TFRecordDataset | TensorFlow Core v2.6.0]
• [tf.data: Build TensorFlow input pipelines | TensorFlow Core]
• [Vertex AI Training | Google Cloud]
• [NVIDIA Tesla V100 GPU | NVIDIA]

According to the official exam guide1, one of the skills assessed in the exam is to “design, build, and productionalize ML models to solve business challenges using Google Cloud technologies”. Document AI2 is a document understanding platform that takes unstructured data from documents and transforms it into structured data, making it easier to understand, analyze, and consume. Document AI Workbench3 allows you to create custom extractors to parse the text in specific sections of your documents. Natural Language API4 is a service that provides natural language understanding technologies, such as sentiment analysis, entity analysis, and other text annotations. The analyzeSentiment feature5 inspects the given text and identifies the prevailing emotional opinion within the text, especially to determine a writer’s attitude as positive, negative, or neutral. Therefore, option C is the best way to accomplish the task of predicting an overall satisfaction score from the customer comments on each form. The other options are not relevant or optimal for this scenario. References:

• Professional ML Engineer Exam Guide
• Document AI
• Document AI Workbench
• Natural Language API
• Sentiment analysis
• Google Professional Machine Learning Certification Exam 2023
• Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

According to the web search results, Vertex AI is a unified platform for machine learning development and deployment. Vertex AI offers various services and tools for building, managing, and serving machine learning models1. Vertex AI allows you to deploy your models to endpoints for online prediction, and configure the compute resources and autoscaling options for your deployed models2. Autoscaling with Vertex AI endpoints is (by default) based on the CPU utilization across all cores of the machine type you have specified. The default threshold of 60% represents 60% on all cores. For example, for a 4 core machine, that means you need 240% utilization to trigger autoscaling3. Therefore, if you discover that the endpoint does not autoscale as expected when receiving multiple requests, you might need to decrease the CPU utilization target in the autoscaling configurations. This way, you can lower the threshold for triggering autoscaling and allocate more resources to handle the prediction requests. Therefore, option D is the best way to solve the problem for the given use case. The other options are not relevant or optimal for this scenario. References:

• Vertex AI
• Deploy a model to an endpoint
• Vertex AI endpoint doesn’t scale up / down
• Google Professional Machine Learning Certification Exam 2023
• Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

 The best option for developing an image classification model by using a large dataset that contains labeled images in a Cloud Storage bucket is to import the labeled images as a managed dataset in Vertex AI and use AutoML to train the model. This option allows you to leverage the power and simplicity of Google Cloud to create and deploy a high-quality image classification model with minimal code and configuration. Vertex AI is a unified platform for building and deploying machine learning solutions on Google Cloud. Vertex AI can create a managed dataset from a Cloud Storage bucket that contains labeled images, which can be used to train an AutoML model. AutoML is a service that can automatically build and optimize machine learning models for various tasks, such as image classification, object detection, natural language processing, and tabular data analysis. AutoML can handle the complex aspects of machine learning, such as feature engineering, model architecture, hyperparameter tuning, and model evaluation. AutoML can also evaluate, deploy, and monitor the image classification model, and provide online or batch predictions. By using Vertex AI and AutoML, users can develop an image classification model by using a large dataset with ease and efficiency.

The other options are not as good as option C, for the following reasons:

• Option A: Using Vertex AI Pipelines with the Kubeflow Pipelines SDK to create a pipeline that reads the images from Cloud Storage and trains the model would require more skills and steps than using Vertex AI and AutoML. Vertex AI Pipelines is a service that can orchestrate machine learning workflows using Vertex AI. Vertex AI Pipelines can run preprocessing and training steps on custom Docker images, and evaluate, deploy, and monitor the machine learning model. Kubeflow Pipelines SDK is a Python library that can create and run pipelines on Vertex AI Pipelines or on Kubeflow, an open-source platform for machine learning on Kubernetes. However, using Vertex AI Pipelines and Kubeflow Pipelines SDK would require writing code, building Docker images, defining pipeline components and steps, and managing the pipeline execution and artifacts. Moreover, Vertex AI Pipelines and Kubeflow Pipelines SDK are not specialized for image classification, and users would need to use other libraries or frameworks, such as TensorFlow or PyTorch, to build and train the image classification model.
• Option B: Using Vertex AI Pipelines with TensorFlow Extended (TFX) to create a pipeline that reads the images from Cloud Storage and trains the model would require more skills and steps than using Vertex AI and AutoML. TensorFlow Extended (TFX) is a framework that can create and run end-to-end machine learning pipelines on TensorFlow, a popular library for building and training deep learning models. TFX can preprocess the data, train and evaluate the model, validate and push the model, and serve the model for online or batch predictions. However, using Vertex AI Pipelines and TFX would require writing code, building Docker images, defining pipeline components and steps, and managing the pipeline execution and artifacts. Moreover, TFX is not optimized for image classification, and users would need to use other libraries or tools, such as TensorFlow Data Validation, TensorFlow Transform, and TensorFlow Hub, to handle the image data and the model architecture.
• Option D: Converting the image dataset to a tabular format using Dataflow, loading the data into BigQuery, and using BigQuery ML to train the model would not handle the image data properly and could result in a poor model performance. Dataflow is a service that can create scalable and reliable pipelines to process large volumes of data from various sources. Dataflow can preprocess the data by using Apache Beam, a programming model for defining and executing data processing workflows. BigQuery is a serverless, scalable, and cost-effective data warehouse that can perform fast and interactive queries on large datasets. BigQuery ML is a service that can create and train machine learning models by using SQL queries on BigQuery. However, converting the image data to a tabular format would lose the spatial and semantic information of the images, which are essential for image classification. Moreover, BigQuery ML is not specialized for image classification, and users would need to use other tools or techniques, such as feature hashing, embedding, or one-hot encoding, to handle the categorical features.

• Option A is incorrect because it does not separate the training dataset into two tables based on the features, which is necessary to train the two models separately and accurately.
• Option B is incorrect because it does not separate the training dataset into two tables based on the features, and because it uses the same monitoring-config-from parameter for both models, which would not account for the different feature selections.
• Option C is incorrect because it deploys the models to two separate endpoints, which would increase the management effort and complexity of the pipeline.
• Option D is correct because it separates the training dataset into two tables based on the features, which would enable the two models to be trained separately and accurately. It also deploys both models to the same endpoint, which would simplify the pipeline and reduce the management effort. It also submits a Vertex Al Model Monitoring job with a monitoring-config-from parameter that accounts for the model IDs and training datasets, which would enable the skew detection to work properly for each model.

The class imbalance problem is a common challenge in machine learning, especially in classification tasks. It occurs when the distribution of the target classes is highly skewed, such that one class (the majority class) has much more examples than the other class (the minority class). The minority class is often the more interesting or important class, such as failure incidents, fraud cases, or rare diseases. However, most machine learning algorithms are designed to optimize the overall accuracy, which can be biased towards the majority class and ignore the minority class. This can result in poor predictive performance, especially for the minority class.

There are different techniques to deal with the class imbalance problem, such as data-level methods, algorithm-level methods, and evaluation-level methods1. Data-level methods involve resampling the original dataset to create a more balanced class distribution. There are two main types of data-level methods: oversampling and undersampling. Oversampling methods increase the number of examples in the minority class, either by duplicating existing examples or by generating synthetic examples. Undersampling methods reduce the number of examples in the majority class, either by randomly removing examples or by using clustering or other criteria to select representative examples. Both oversampling and undersampling methods can be combined with upweighting or downweighting, which assign different weights to the examples according to their class frequency, to further balance the dataset.

For the use case of investigating failures of a production line component based on sensor readings, the best option is to downsample the data with upweighting to create a sample with 10% positive examples. This option involves randomly removing some of the negative examples (the majority class) until the ratio of positive to negative examples is 1:9, and then assigning higher weights to the positive examples to compensate for their low frequency. This option can create a more balanced dataset that can improve the performance of the classification models, while preserving the diversity and representativeness of the original data. This option can also reduce the computation time and memory usage, as the size of the dataset is reduced. Therefore, downsampling the data with upweighting to create a sample with 10% positive examples is the best option for this use case.

References:

• A Systematic Study of the Class Imbalance Problem in Convolutional Neural Networks

• Option A is incorrect because using AI Platform notebooks to perform a Lasso regression analysis on your model, which will eliminate features that do not provide a strong signal, is not a suitable way to determine which customer attribute has the most predictive power for each prediction served by the model. Lasso regression is a method of feature selection that applies a penalty to the coefficients of the linear model, and shrinks them to zero for irrelevant features1. However, this method assumes that the model is linear and additive, which may not be the case for a TensorFlow model. Moreover, this method does not provide feature attributions for each prediction, but rather for the entire dataset.
• Option B is incorrect because streaming prediction results to BigQuery, and using BigQuery’s CORR(X1, X2) function to calculate the Pearson correlation coefficient between each feature and the target variable, is not a valid way to determine which customer attribute has the most predictive power for each prediction served by the model. The Pearson correlation coefficient is a measure of the linear relationship between two variables, ranging from -1 to 12. However, this method does not account for the interactions between features or the non-linearity of the model. Moreover, this method does not provide feature attributions for each prediction, but rather for the entire dataset.
• Option C is correct because using the AI Explanations feature on AI Platform, and submitting each prediction request with the ‘explain’ keyword to retrieve feature attributions using the sampled Shapley method, is the best way to determine which customer attribute has the most predictive power for each prediction served by the model. AI Explanations is a service that allows you to get feature attributions for your deployed models on AI Platform3. Feature attributions are values that indicate how much each feature contributed to the prediction for a given instance4. The sampled Shapley method is a technique that uses the Shapley value, a game-theoretic concept, to measure the contribution of each feature to the prediction5. By using AI Explanations, you can get feature attributions for each prediction request, and identify the most important features for each customer.
• Option D is incorrect because using the What-If tool in Google Cloud to determine how your model will perform when individual features are excluded, and ranking the feature importance in order of those that caused the most significant performance drop when removed from the model, is not a practical way to determine which customer attribute has the most predictive power for each prediction served by the model. The What-If tool is a tool that allows you to visualize and analyze your ML models and datasets. However, this method requires manually editing or removing features for each instance, and observing the change in the prediction. This method is not scalable or efficient, and may not capture the interactions between features or the non-linearity of the model.

References:

• Lasso regression
• Pearson correlation coefficient
• AI Explanations overview
• Feature attributions
• Sampled Shapley method
• [What-If tool overview]

Batch prediction is the process of using an ML model to make predictions on a large set of data points. Batch prediction is suitable for scenarios where the predictions are not time-sensitive and can be done in batches, such as digitizing scanned customer forms at the end of each day. Batch prediction can also handle large volumes of data and scale up or down the resources as needed. AI Platform provides a batch prediction service that allows users to submit a job with their TensorFlow model and input data stored in Cloud Storage, and receive the output predictions in Cloud Storage as well. This service requires minimal manual intervention and can be automated with Cloud Scheduler or Cloud Functions. Therefore, using the batch prediction functionality of AI Platform is the best option for this use case.

References:

• Batch prediction overview
• Using batch prediction

 Overfitting occurs when a model tries to fit the training data so closely that it does not generalize well to new data. Overfitting can be caused by having a model that is too complex for the data, such as having too many parameters or layers. Overfitting can lead to poor performance on the validation data, which reflects how the model will perform on unseen data1

To prevent overfitting, one strategy is to use regularization techniques that penalize the complexity of the model and encourage it to learn simpler patterns. Two common regularization techniques for deep neural networks are L2 regularization and dropout. L2 regularization adds a term to the loss function that is proportional to the squared magnitude of the model’s weights. This term penalizes large weights and encourages the model to use smaller weights. Dropout randomly drops out some units in the network during training, which prevents co-adaptation of features and reduces the effective number of parameters. Both L2 regularization and dropout have hyperparameters that control the strength of the regularization effect23

Another strategy to prevent overfitting is to use hyperparameter tuning, which is the process of finding the optimal values for the parameters of the model that affect its performance. Hyperparameter tuning can help find the best combination of hyperparameters that minimize the validation loss and improve the generalization ability of the model. AI Platform provides a service for hyperparameter tuning that can run multiple trials in parallel and use different search algorithms to find the best solution.

Therefore, the best strategy to use when retraining the model is to run a hyperparameter tuning job on AI Platform to optimize for the L2 regularization and dropout parameters. This will allow the model to find the optimal balance between fitting the training data and generalizing to new data. The other options are not as effective, as they either use fixed values for the regularization parameters, which may not be optimal, or they do not address the issue of overfitting at all.

References: 1: Generalization: Peril of Overfitting 2: Regularization for Deep Learning 3: Dropout: A Simple Way to Prevent Neural Networks from Overfitting : [Hyperparameter tuning overview]

 This answer is correct because it uses Firebase, a platform that provides a scalable and reliable notification system for mobile and web applications. Firebase Cloud Messaging (FCM) allows you to send messages and notifications to users across different devices and platforms. By registering each user with a user ID on the FCM server, you can target specific users based on their account balance predictions and send them personalized notifications when their balance is likely to drop below the $25 threshold. This way, you can provide a useful and timely feature for your customers and increase their engagement and retention. References:

• [Firebase Cloud Messaging]
• [Firebase Cloud Messaging: Send messages to specific devices]

 Sentiment analysis is the process of identifying and extracting the emotions, opinions, and attitudes expressed in a text or speech. Sentiment analysis can help businesses understand their customers’ feedback, satisfaction, and preferences. There are different approaches to building a sentiment analysis tool, depending on the input data and the output format. Some of the common approaches are:

• Extracting sentiment directly from the voice recordings: This approach involves using acoustic features, such as pitch, intensity, and prosody, to infer the sentiment of the speaker. This approach can capture the nuances and subtleties of the vocal expression, but it also requires a large and diverse dataset of labeled voice recordings, which may not be easily available or accessible. Moreover, this approach may not account for the semantic and contextual information of the speech, which can also affect the sentiment.
• Converting the speech to text and building a model based on the words: This approach involves using automatic speech recognition (ASR) to transcribe the voice recordings into text, and then using lexical features, such as word frequency, polarity, and valence, to infer the sentiment of the text. This approach can leverage the existing text-based sentiment analysis models and tools, but it also introduces some challenges, such as the accuracy and reliability of the ASR system, the ambiguity and variability of the natural language, and the loss of the acoustic information of the speech.
• Converting the speech to text and extracting sentiments based on the sentences: This approach involves using ASR to transcribe the voice recordings into text, and then using syntactic and semantic features, such as sentence structure, word order, and meaning, to infer the sentiment of the text. This approach can capture the higher-level and complex aspects of the natural language, such as negation, sarcasm, and irony, which can affect the sentiment. However, this approach also requires more sophisticated and advanced natural language processing techniques, such as parsing, dependency analysis, and semantic role labeling, which may not be readily available or easy to implement.
• Converting the speech to text and extracting sentiment using syntactical analysis: This approach involves using ASR to transcribe the voice recordings into text, and then using syntactical analysis, such as part-of-speech tagging, phrase chunking, and constituency parsing, to infer the sentiment of the text. This approach can identify the grammatical and structural elements of the natural language, such as nouns, verbs, adjectives, and clauses, which can indicate the sentiment. However, this approach may not account for the pragmatic and contextual information of the speech, such as the speaker’s intention, tone, and situation, which can also influence the sentiment.

For the use case of building a sentiment analysis tool that predicts customer sentiment from recorded phone conversations, the best approach is to convert the speech to text and extract sentiments based on the sentences. This approach can balance the trade-offs between the accuracy, complexity, and feasibility of the sentiment analysis tool, while ensuring that the gender, age, and cultural differences of the customers who called the contact center do not impact any stage of the model development pipeline and results. This approach can also handle different types and levels of sentiment, such as polarity (positive, negative, or neutral), intensity (strong or weak), and emotion (anger, joy, sadness, etc.). Therefore, converting the speech to text and extracting sentiments based on the sentences is the best approach for this use case.

The best option to stabilize the pipeline without downgrading the evaluation quality while minimizing infrastructure overhead is to use Dataflow as the runner for the evaluation step. Dataflow is a fully managed service for executing Apache Beam pipelines that can scale up and down according to the workload. Dataflow can handle large-scale, distributed data processing tasks such as model evaluation, and it can also integrate with Vertex AI Pipelines and TensorFlow Extended (TFX). By using the flag -runner=DataflowRunner in beam_pipeline_args, you can instruct the Evaluator component to run the evaluation step on Dataflow, instead of using the default DirectRunner, which runs locally and may cause out-of-memory errors. Option A is incorrect because adding tfma.MetricsSpec() to limit the number of metrics in the evaluation step may downgrade the evaluation quality, as some important metrics may be omitted. Moreover, reducing the number of metrics may not solve the out-of-memory error, as the evaluation step may still consume a lot of memory depending on the size and complexity of the data and the model. Option B is incorrect because migrating the pipeline to Kubeflow hosted on Google Kubernetes Engine (GKE) may increase the infrastructure overhead, as you need to provision, manage, and monitor the GKE cluster yourself. Moreover, you need to specify the appropriate node parameters for the evaluation step, which may require trial and error to find the optimal configuration. Option D is incorrect because moving the evaluation step out of the pipeline and running it on custom Compute Engine VMs may also increase the infrastructure overhead, as you need to create, configure, and delete the VMs yourself. Moreover, you need to ensure that the VMs have sufficient memory for the evaluation step, which may require trial and error to find the optimal machine type. References:

• Dataflow documentation
• Using DataflowRunner
• Evaluator component documentation
• Configuring the Evaluator component

 TFX (TensorFlow Extended) is a platform for end-to-end machine learning pipelines. It provides components for data ingestion, preprocessing, validation, model training, serving, and monitoring. Dataflow is a fully managed service for scalable data processing. By using TFX components with Dataflow, you can perform feature engineering on large-scale tabular data in a distributed and efficient way. You can use the Transform component to apply the MaxMin scaler and the one-hot encoding to the numerical and categorical features, respectively. You can also use the ExampleGen component to read data from BigQuery and the Trainer component to train your TensorFlow model. The output of the Transform component is a TFRecord file, which is a binary format for storing TensorFlow data. You can export the TFRecord file to Cloud Storage and feed it into Vertex AI Training, which is a managed service for training custom machine learning models on Google Cloud. References:

• TFX | TensorFlow
• Dataflow | Google Cloud
• Vertex AI Training | Google Cloud

 The best option for accessing internal data in the most secure way, while mitigating the risk of data exfiltration, is to enable VPC Service Controls for peerings, and add Vertex AI to a service perimeter. This option allows you to leverage the power and simplicity of VPC Service Controls to isolate and protect your data and services on Google Cloud. VPC Service Controls is a service that can create a secure perimeter around your Google Cloud resources, such as BigQuery, Cloud Storage, and Vertex AI. VPC Service Controls can help you prevent unauthorized access and data exfiltration from your perimeter, and enforce fine-grained access policies based on context and identity. Peerings are connections that can allow traffic to flow between different networks. Peerings can help you connect your Google Cloud network with other Google Cloud networks or external networks, and enable communication between your resources and services. By enabling VPC Service Controls for peerings, you can allow your training code to download internal data by using an API endpoint hosted in your project’s network, and restrict the data transfer to only authorized networks and services. Vertex AI is a unified platform for building and deploying machine learning solutions on Google Cloud. Vertex AI can support various types of models, such as linear regression, logistic regression, k-means clustering, matrix factorization, and deep neural networks. Vertex AI can also provide various tools and services for data analysis, model development, model deployment, model monitoring, and model governance. By adding Vertex AI to a service perimeter, you can isolate and protect your Vertex AI resources, such as models, endpoints, pipelines, and feature store, and prevent data exfiltration from your perimeter1.

The other options are not as good as option A, for the following reasons:

• Option B: Creating a Cloud Run endpoint as a proxy to the data, and using Identity and Access Management (IAM) authentication to secure access to the endpoint from the training job would require more skills and steps than enabling VPC Service Controls for peerings, and adding Vertex AI to a service perimeter. Cloud Run is a service that can run your stateless containers on a fully managed environment or on your own Google Kubernetes Engine cluster. Cloud Run can help you deploy and scale your containerized applications quickly and easily, and pay only for the resources you use. A Cloud Run endpoint is a URL that can expose your containerized application to the internet or to other Google Cloud services. A Cloud Run endpoint can help you access and invoke your application from anywhere, and handle the load balancing and traffic routing. A proxy is a server that can act as an intermediary between a client and a target server. A proxy can help you modify, filter, or redirect the requests and responses between the client and the target server, and provide additional functionality or security. IAM is a service that can manage access control for Google Cloud resources. IAM can help you define who (identity) has what access (role) to which resource, and enforce the access policies. By creating a Cloud Run endpoint as a proxy to the data, and using IAM authentication to secure access to the endpoint from the training job, you can access internal data by using an API endpoint hosted in your project’s network, and restrict the data access to only authorized identities and roles. However, creating a Cloud Run endpoint as a proxy to the data, and using IAM authentication to secure access to the endpoint from the training job would require more skills and steps than enabling VPC Service Controls for peerings, and adding Vertex AI to a service perimeter. You would need to write code, create and configure the Cloud Run endpoint, implement the proxy logic, deploy and monitor the Cloud Run endpoint, and set up the IAM policies. Moreover, this option would not prevent data exfiltration from your network, as the Cloud Run endpoint can be accessed from outside your network2.
• Option C: Configuring VPC Peering with Vertex AI and specifying the network of the training job would not allow you to access internal data by using an API endpoint hosted in your project’s network, and could cause errors or poor performance. VPC Peering is a service that can create a peering connection between two VPC networks. VPC Peering can help you connect your Google Cloud network with another Google Cloud network or an external network, and enable communication between your resources and services. By configuring VPC Peering with Vertex AI and specifying the network of the training job, you can allow your training code to access Vertex AI resources, such as models, endpoints, pipelines, and feature store, and use the same network for the training job. However, configuring VPC Peering with Vertex AI and specifying the network of the training job would not allow you to access internal data by using an API endpoint hosted in your project’s network, and could cause errors or poor performance. You would need to write code, create and configure the VPC Peering connection, and specify the network of the training job. Moreover, this option would not isolate and protect your data and services on Google Cloud, as the VPC Peering connection can expose your network to other networks and services3.
• Option D: Downloading the data to a Cloud Storage bucket before calling the training job would not allow you to access internal data by using an API endpoint hosted in your project’s network, and could increase the complexity and cost of the data access. Cloud Storage is a service that can store and manage your data on Google Cloud. Cloud Storage can help you upload and organize your data, and track the data versions and metadata. A Cloud Storage bucket is a container that can hold your data on Cloud Storage. A Cloud Storage bucket can help you store and access your data from anywhere, and provide various storage classes and options. By downloading the data to a Cloud Storage bucket before calling the training job, you can access the data from Cloud Storage, and use it as the input for the training job. However, downloading the data to a Cloud Storage bucket before calling the training job would not allow you to access internal data by using an API endpoint hosted in your project’s network, and could increase the complexity and cost of the data access. You would need to write code, create and configure the Cloud Storage bucket, download the data to the Cloud Storage bucket, and call the training job. Moreover, this option would create an intermediate data source on Cloud Storage, which can increase the storage and transfer costs, and expose the data to unauthorized access or data exfiltration4.

References:

• Preparing for Google Cloud Certification: Machine Learning Engineer, Course 3: Production ML Systems, Week 1: Data Engineering
• Google Cloud Professional Machine Learning Engineer Exam Guide, Section 1: Framing ML problems, 1.2 Defining data needs
• Official Google Cloud Certified Professional Machine Learning Engineer Study Guide, Chapter 2: Data Engineering, Section 2.2: Defining Data Needs
• VPC Service Controls
• Cloud Run
• VPC Peering
• Cloud Storage

• Option A is incorrect because training a time-series model to predict the machines’ performance values, and configuring an alert if a machine’s actual performance values significantly differ from the predicted performance values, is not the best way to build a predictive maintenance solution that uses monitoring data from the VMs to detect potential failures and then alerts the service desk team. This option assumes that the performance values follow a predictable pattern, which may not be the case for complex systems. Moreover, this option does not use any historical incident data, which may contain useful information for identifying failures. Furthermore, this option does not involve any model evaluation or validation, which are essential steps for ensuring the quality and reliability of the model.
• Option B is correct because implementing a simple heuristic (e.g., based on z-score) to label the machines’ historical performance data, and training a model to predict anomalies based on this labeled dataset, is a reasonable way to build a predictive maintenance solution that uses monitoring data from the VMs to detect potential failures and then alerts the service desk team. This option uses a simple and fast method to label the historical performance data, which is necessary for supervised learning. A z-score is a measure of how many standard deviations a value is away from the mean of a distribution1. By using a z-score, we can label the performance values that are unusually high or low as anomalies, which may indicate failures. Then, we can train a model to learn the patterns of normal and anomalous performance values, and use it to predict anomalies on new data. We can also evaluate and validate the model using metrics such as precision, recall, or F1-score, and compare it with other models or methods.
• Option C is incorrect because developing a simple heuristic (e.g., based on z-score) to label the machines’ historical performance data, and testing this heuristic in a production environment, is not a safe way to build a predictive maintenance solution that uses monitoring data from the VMs to detect potential failures and then alerts the service desk team. This option does not involve any model training or evaluation, which are essential steps for ensuring the quality and reliability of the solution. Moreover, this option does not test the heuristic on a separate dataset, such as a validation or test set, before deploying it to production, which may lead to errors or failures in the production environment.
• Option D is incorrect because hiring a team of qualified analysts to review and label the machines’ historical performance data, and training a model based on this manually labeled dataset, is not a feasible way to build a predictive maintenance solution that uses monitoring data from the VMs to detect potential failures and then alerts the service desk team. This option may produce high-quality labels, but it is also costly, time-consuming, and prone to human errors or biases. Moreover, this option may not scale well with large or complex datasets, which may require more analysts or more time to label.

References:

• Z-score
• [Predictive maintenance]
• [Anomaly detection]
• [Time-series analysis]
• [Model evaluation]

• Option A is incorrect because creating a one-hot encoding of words, and feeding the encodings into your model is not an efficient way to preprocess the words individually for a natural language model. One-hot encoding is a method of representing categorical variables as binary vectors, where each element corresponds to a category and only one element is 1 and the rest are 01. However, this method is not suitable for high-dimensional and sparse data, such as words in a large vocabulary, because it requires a lot of memory and computation, and does not capture the semantic similarity or relationship between words2.
• Option B is correct because identifying word embeddings from a pre-trained model, and using the embeddings in your model is a good way to preprocess the words individually for a natural language model. Word embeddings are low-dimensional and dense vectors that represent the meaning and usage of words in a continuous space3. Word embeddings can be learned from a large corpus of text using neural networks, such as word2vec, GloVe, or BERT4. Using pre-trained word embeddings can save time and resources, and improve the performance of the natural language model, especially when the training data is limited or noisy5.
• Option C is incorrect because sorting the words by frequency of occurrence, and using the frequencies as the encodings in your model is not a meaningful way to preprocess the words individually for a natural language model. This method implies that the frequency of a word is a good indicator of its importance or relevance, which may not be true. For example, the word “the” is very frequent but not very informative, while the word “unicorn” is rare but more distinctive. Moreover, this method does not capture the semantic similarity or relationship between words, and may introduce noise or bias into the model.
• Option D is incorrect because assigning a numerical value to each word from 1 to 100,000 and feeding the values as inputs in your model is not a valid way to preprocess the words individually for a natural language model. This method implies an ordinal relationship between the words, which may not be true. For example, assigning the values 1, 2, and 3 to the words “apple”, “banana”, and “orange” does not make sense, as there is no inherent order among these fruits. Moreover, this method does not capture the semantic similarity or relationship between words, and may confuse the model with irrelevant or misleading information.

References:

• One-hot encoding
• Word embeddings
• Word embedding
• Pre-trained word embeddings
• Using pre-trained word embeddings in a Keras model
• [Term frequency]
• [Term frequency-inverse document frequency]
• [Ordinal variable]
• [Encoding categorical features]

The best option for productionizing a Keras model is to use TensorFlow Extended (TFX), a framework for building end-to-end machine learning pipelines that can handle large-scale data and complex workflows. TFX provides standard components for data ingestion, transformation, validation, analysis, training, tuning, serving, and monitoring. TFX pipelines can be orchestrated with Vertex AI Pipelines, a managed service that runs on Google Cloud Platform and leverages Kubernetes and Argo. Vertex AI Pipelines allows you to automate the execution of your TFX pipeline steps, schedule retraining jobs, and scale up or down the resources as needed. By using TFX and Vertex AI Pipelines, you can take advantage of the following benefits:

• You can reuse the existing code in your Jupyter notebook, as TFX supports Keras as a first-class citizen. You can also use the Keras Tuner to optimize your model hyperparameters.
• You can ensure data quality and consistency by using the TFX Data Validation component, which can detect anomalies, drift, and skew in your data. You can also use the TFX SchemaGen component to generate a schema for your data and enforce it throughout the pipeline.
• You can analyze your model performance and fairness by using the TFX Model Analysis component, which can produce various metrics and visualizations. You can also use the TFX Model Validation component to compare your new model with a baseline model and set thresholds for deploying the model to production.
• You can deploy your model to various serving platforms by using the TFX Pusher component, which can push your model to Vertex AI, Cloud AI Platform, TensorFlow Serving, or TensorFlow Lite. You can also use the TFX Model Registry to manage the versions and metadata of your models.
• You can monitor your model performance and health by using the TFX Model Monitor component, which can detect data drift, concept drift, and prediction skew in your model. You can also use the TFX Evaluator component to compute metrics and validate your model against a baseline or a slice of data.
• You can reduce the cost and complexity of managing your own infrastructure by using Vertex AI Pipelines, which provides a serverless environment for running your TFX pipeline. You can also use the Vertex AI Experiments and Vertex AI TensorBoard to track and visualize your pipeline runs.

References:

• [TensorFlow Extended (TFX)]
• [Vertex AI Pipelines]
• [TFX User Guide]

This is not a good result because the model is performing worse than predicting that people will always renew their subscription. This option has the following reasons:

• It indicates that the model is not learning from the data, but rather memorizing the majority class. Since 90% of the individuals renew their subscription every year, the model can achieve a 90% accuracy by simply predicting that everyone will renew their subscription, without considering the features or the patterns in the data. However, the model’s accuracy for predicting those who renew their subscription is only 82%, which is lower than the baseline accuracy of 90%. This suggests that the model is overfitting to the minority class (those who cancel their subscription), and underfitting to the majority class (those who renew their subscription).
• It implies that the model is not useful for the business problem, as it cannot identify the customers who are at risk of churning. The goal of predicting whether customers will cancel their annual subscription is to prevent customer churn and increase customer retention. However, the model’s accuracy for predicting those who cancel their subscription is 99%, which is too high and unrealistic, as it means that the model can almost perfectly identify the customers who will churn, without any false positives or false negatives. This may indicate that the model is cheating or exploiting some leakage in the data, such as a feature that reveals the outcome of the prediction. Moreover, the model’s accuracy for predicting those who renew their subscription is 82%, which is too low and unreliable, as it means that the model can miss many customers who will churn, and falsely label them as renewing customers. This can lead to losing customers and revenue, and failing to take proactive actions to retain them.

References:

• How to Evaluate Machine Learning Models: Classification Metrics | Machine Learning Mastery
• Imbalanced Classification: Predicting Subscription Churn | Machine Learning Mastery

• Option A is incorrect because migrating your model to TensorFlow, and training it using Vertex AI Training, is not the easiest way to improve the model’s training time. TensorFlow is a framework that allows you to create and train ML models using Python or other languages. Vertex AI Training is a service that allows you to train and optimize ML models using built-in algorithms or custom containers. However, this option requires significant code changes, as TensorFlow and scikit-learn have different APIs and functionalities. Moreover, this option does not leverage the parallelism or the scalability of the cloud, as it only uses a single instance.
• Option B is incorrect because training your model in a distributed mode using multiple Compute Engine VMs, is not the most convenient way to improve the model’s training time. Compute Engine is a service that allows you to create and manage virtual machines that run on Google Cloud. You can use Compute Engine to run your scikit-learn model in a distributed mode, by using libraries such as Dask or Joblib. However, this option requires more effort and resources than option D, as it involves creating and configuring the VMs, installing and maintaining the libraries, and writing and running the distributed code.
• Option C is incorrect because training your model with DLVM images on Vertex AI, and ensuring that your code utilizes NumPy and SciPy internal methods whenever possible, is not the most effective way to improve the model’s training time. DLVM (Deep Learning Virtual Machine) images are preconfigured VM images that include popular ML frameworks and tools, such as TensorFlow, PyTorch, or scikit-learn1. You can use DLVM images on Vertex AI to train your scikit-learn model, by using a custom container. NumPy and SciPy are libraries that provide numerical and scientific computing functionalities for Python. You can use NumPy and SciPy internal methods to optimize your scikit-learn code, as they are faster and more efficient than pure Python code2. However, this option does not leverage the parallelism or the scalability of the cloud, as it only uses a single instance. Moreover, this option may not have a significant impact on the training time, as scikit-learn already relies on NumPy and SciPy for most of its operations3.
• Option D is correct because training your model using Vertex AI Training with GPUs, is the best way to improve the model’s training time. A GPU (Graphics Processing Unit) is a hardware accelerator that can perform parallel computations faster than a CPU (Central Processing Unit)4. Vertex AI Training is a service that allows you to train and optimize ML models using built-in algorithms or custom containers. You can use Vertex AI Training with GPUs to train your scikit-learn model, by using a custom container and specifying the accelerator type and count5. By using Vertex AI Training with GPUs, you can leverage the parallelism and the scalability of the cloud, and speed up the training process significantly, without changing your code.

References:

• DLVM images
• NumPy and SciPy
• scikit-learn dependencies
• GPU overview
• Vertex AI Training with GPUs
• [scikit-learn overview]
• [TensorFlow overview]
• [Compute Engine overview]
• [Dask overview]
• [Joblib overview]
• [Vertex AI Training overview]

• Precision measures the fraction of messages flagged by the model that are actually inappropriate, while recall measures the fraction of inappropriate messages that are flagged by the model. These metrics are useful for evaluating how well the model can identify and filter out inappropriate comments.
• Option A is not a good metric because it does not account for the accuracy of the model. The model might flag many messages that are not inappropriate, or miss many messages that are inappropriate.
• Option B is better than option A, but it still does not account for the recall of the model. The model might flag only a few messages that are highly likely to be inappropriate, but miss many other messages that are less obvious but still inappropriate.
• Option C is not a good metric because it does not focus on the messages that are flagged by the model. The random sample of 0.1% of raw messages might contain very few inappropriate messages, making the precision and recall estimates unreliable.

• Monitoring model performance is an essential part of production readiness, as it allows the team to detect and address any issues that may arise after deployment, such as data drift, model degradation, or errors.
• Other Options:

 This answer is correct because it allows AutoML Tables to handle the time signal in the data and split the data accordingly. This ensures that the model is trained on the historical data and evaluated on the more recent data, which is consistent with the prediction task. AutoML Tables can automatically detect and handle temporal features in the data, such as date, time, and duration. By specifying the Time column, AutoML Tables can also perform time-series forecasting and use the time signal to generate additional features, such as seasonality and trend. References:

• [AutoML Tables: Preparing your training data]
• [AutoML Tables: Time-series forecasting]

One-hot encoding is a technique that converts categorical variables into numerical variables by creating dummy variables for each possible category. Each dummy variable has a value of 1 if the original variable belongs to that category, and 0 otherwise1. One-hot encoding can help linear regression models to capture the effect of different categories on the target variable without imposing any ordinal relationship among them2. Dataprep is a service that allows you to explore, clean, and transform your data for analysis and machine learning. You can use Dataprep to apply one-hot encoding to your city name variable and make each city a column with binary values3. This way, you can prepare your data using the least amount of coding while maintaining the predictive variables. Therefore, using Dataprep to transform the state column using a one-hot encoding method is the best option for this use case.

References:

• One Hot Encoding: A Beginner’s Guide
• One-Hot Encoding in Linear Regression Models
• Dataprep documentation

The best option for reducing pipeline execution time and cost, while also minimizing pipeline changes, is to enable caching for the pipeline job, and disable caching for the model training step. This option allows you to leverage the power and simplicity of Vertex AI Pipelines to reuse the output of the data preprocessing step, and avoid unnecessary recomputation. Vertex AI Pipelines is a service that can orchestrate machine learning workflows using Vertex AI. Vertex AI Pipelines can run preprocessing and training steps on custom Docker images, and evaluate, deploy, and monitor the machine learning model. Caching is a feature of Vertex AI Pipelines that can store and reuse the output of a pipeline step, and skip the execution of the step if the input parameters and the code have not changed. Caching can help you reduce the pipeline execution time and cost, as you do not need to re-run the same step with the same input and code. Caching can also help you minimize the pipeline changes, as you do not need to add or remove any pipeline steps or parameters. By enabling caching for the pipeline job, and disabling caching for the model training step, you can create a Vertex AI pipeline that includes two steps. The first step preprocesses 10 TB data, completes in about 1 hour, and saves the result in a Cloud Storage bucket. The second step uses the processed data to train a model. You can update the model’s code to allow you to test different algorithms, and run the pipeline job with caching enabled. The pipeline job will reuse the output of the data preprocessing step from the cache, and skip the execution of the step. The pipeline job will run the model training step with the updated code, and disable the caching for the step. This way, you can reduce the pipeline execution time and cost, while also minimizing pipeline changes1.

The other options are not as good as option D, for the following reasons:

• Option A: Adding a pipeline parameter and an additional pipeline step, depending on the parameter value, the pipeline step conducts or skips data preprocessing and starts model training, would require more skills and steps than enabling caching for the pipeline job, and disabling caching for the model training step. A pipeline parameter is a variable that can be used to control the input or output of a pipeline step. A pipeline parameter can help you customize the pipeline logic and behavior, and experiment with different values. An additional pipeline step is a new instance of a pipeline component that can perform a part of the pipeline workflow, such as data preprocessing or model training. An additional pipeline step can help you extend the pipeline functionality and complexity, and handle different scenarios. However, adding a pipeline parameter and an additional pipeline step, depending on the parameter value, the pipeline step conducts or skips data preprocessing and starts model training, would require more skills and steps than enabling caching for the pipeline job, and disabling caching for the model training step. You would need to write code, define the pipeline parameter, create the additional pipeline step, implement the conditional logic, and compile and run the pipeline. Moreover, this option would not reuse the output of the data preprocessing step from the cache, but rather from the Cloud Storage bucket, which can increase the data transfer and access costs1.
• Option B: Creating another pipeline without the preprocessing step, and hardcoding the preprocessed Cloud Storage file location for model training, would require more skills and steps than enabling caching for the pipeline job, and disabling caching for the model training step. A pipeline without the preprocessing step is a pipeline that only includes the model training step, and uses the preprocessed data from the Cloud Storage bucket as the input. A pipeline without the preprocessing step can help you avoid running the data preprocessing step every time, and reduce the pipeline execution time and cost. However, creating another pipeline without the preprocessing step, and hardcoding the preprocessed Cloud Storage file location for model training, would require more skills and steps than enabling caching for the pipeline job, and disabling caching for the model training step. You would need to write code, create a new pipeline, remove the preprocessing step, hardcode the Cloud Storage file location, and compile and run the pipeline. Moreover, this option would not reuse the output of the data preprocessing step from the cache, but rather from the Cloud Storage bucket, which can increase the data transfer and access costs. Furthermore, this option would create another pipeline, which can increase the maintenance and management costs1.
• Option C: Configuring a machine with more CPU and RAM from the compute-optimized machine family for the data preprocessing step, would not reduce the pipeline execution time and cost, while also minimizing pipeline changes, but rather increase the pipeline execution cost and complexity. A machine with more CPU and RAM from the compute-optimized machine family is a virtual machine that has a high ratio of CPU cores to memory, and can provide high performance and scalability for compute-intensive workloads. A machine with more CPU and RAM from the compute-optimized machine family can help you optimize the data preprocessing step, and reduce the pipeline execution time. However, configuring a machine with more CPU and RAM from the compute-optimized machine family for the data preprocessing step, would not reduce the pipeline execution time and cost, while also minimizing pipeline changes, but rather increase the pipeline execution cost and complexity. You would need to write code, configure the machine type parameters for the data preprocessing step, and compile and run the pipeline. Moreover, this option would increase the pipeline execution cost, as machines with more CPU and RAM from the compute-optimized machine family are more expensive than machines with less CPU and RAM from other machine families. Furthermore, this option would not reuse the output of the data preprocessing step from the cache, but rather re-run the data preprocessing step every time, which can increase the pipeline execution time and cost1.

References:

• Preparing for Google Cloud Certification: Machine Learning Engineer, Course 3: Production ML Systems, Week 3: MLOps
• Google Cloud Professional Machine Learning Engineer Exam Guide, Section 3: Scaling ML models in production, 3.2 Automating ML workflows
• Official Google Cloud Certified Professional Machine Learning Engineer Study Guide, Chapter 6: Production ML Systems, Section 6.4: Automating ML Workflows
• Vertex AI Pipelines
• Caching
• Pipeline parameters
• Machine types

• Reasoning: The question asks for a design that serves asynchronous predictions to determine whether a machine part will fail. This means that the predictions do not need to be returned immediately to the sensors, but can be processed in batches and sent to a downstream system for monitoring. Option B is the only one that uses a streaming data pipeline with Pub/Sub and Dataflow, which can handle real-time data ingestion, processing, and prediction. Option B also invokes the model for prediction, which is required by the question. The other options either use synchronous predictions (option A), batch predictions (options C and D), or do not invoke the model for prediction (option D).
• References: You can learn more about the differences between synchronous, asynchronous, and batch predictions in Vertex AI from this document. You can also find examples of how to use Pub/Sub and Dataflow for streaming data pipelines from this tutorial and this codelab.

AutoML Tables is a service that allows you to automatically build and deploy state-of-the-art machine learning models on structured data without writing code. You can use AutoML Tables to perform the following steps for the classification task:

• Exploratory data analysis: AutoML Tables provides a graphical user interface (GUI) and a command-line interface (CLI) to explore your data, visualize statistics, and identify potential issues.
• Feature selection: AutoML Tables automatically selects the most relevant features for your model based on the data schema and the target column. You can also manually exclude or include features, or create new features from existing ones using feature engineering.
• Model building: AutoML Tables automatically builds and evaluates multiple machine learning models using different algorithms and architectures. You can also specify the optimization objective, the budget, and the evaluation metric for your model.
• Training and hyperparameter tuning: AutoML Tables automatically trains and tunes your model using the best practices and techniques from Google’s research and engineering teams. You can monitor the training progress and the performance of your model on the GUI or the CLI.
• Serving: AutoML Tables automatically deploys your model to a fully managed, scalable, and secure environment. You can use the GUI or the CLI to request predictions from your model, either online (synchronously) or offline (asynchronously).

References:

• [AutoML Tables documentation]
• [AutoML Tables overview]
• [AutoML Tables how-to guides]

 Vertex AI Experiments is a service that allows you to track, compare, and optimize your ML experiments on Google Cloud. You can use Vertex AI Experiments to log metrics and parameters from your TensorFlow models, and then visualize them in Vertex AI TensorBoard. Vertex AI TensorBoard is a managed service that provides a web interface for viewing and debugging your ML experiments. You can use Vertex AI TensorBoard to compare different runs, inspect model graphs, analyze scalars, histograms, images, and more. By using Vertex AI Experiments and Vertex AI TensorBoard, you can simplify your ML experiment tracking and visualization workflow, and avoid the overhead of setting up and maintaining your own Cloud Functions, Cloud Storage buckets, or VMs. References:

• [Vertex AI Experiments documentation]
• [Vertex AI TensorBoard documentation]
• Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate

To avoid creating or reinforcing unfair bias in the model, you should collect a representative sample of production traffic to build the training dataset, and conduct fairness tests across sensitive categories and demographics on the trained model. A representative sample is one that reflects the true distribution of the population, and does not over- or under-represent any group. A random sample is a simple way to obtain a representative sample, as it ensures that every data point has an equal chance of being selected. A stratified sample is another way to obtain a representative sample, as it ensures that every subgroup has a proportional representation in the sample. However, a stratified sample requires prior knowledge of the subgroups and their sizes, which may not be available or easy to obtain. Therefore, a random sample is a more feasible option in this case. A fairness test is a way to measure and evaluate the potential bias and discrimination of the model, based on different categories and demographics, such as age, gender, race, etc. A fairness test can help you identify and mitigate any unfair outcomes or impacts of the model, and ensure that the model treats all groups fairly and equitably. A fairness test can be conducted using various methods and tools, such as confusion matrices, ROC curves, fairness indicators, etc. References: The answer can be verified from official Google Cloud documentation and resources related to data sampling and fairness testing.

• Sampling data | BigQuery
• Fairness Indicators | TensorFlow
• What-if Tool | TensorFlow

• Option A is correct because importing the TensorFlow model with BigQuery ML, and running the ml.predict function is the easiest way to execute a batch prediction on a large BigQuery table with a custom TensorFlow model, and store the predicted results in another BigQuery table. BigQuery ML allows you to import TensorFlow models that are stored in Cloud Storage, and use them for prediction with SQL queries1. The ml.predict function returns a table with the predicted values, which can be saved to another BigQuery table2.
• Option B is incorrect because using the TensorFlow BigQuery reader to load the data, and using the BigQuery API to write the results to BigQuery requires more effort to build the inference pipeline than option A. The TensorFlow BigQuery reader is a way to read data from BigQuery into TensorFlow datasets, which can be used for training or prediction3. However, this option also requires writing code to load the TensorFlow model, run the prediction, and use the BigQuery API to write the results back to BigQuery4.
• Option C is incorrect because creating a Dataflow pipeline to convert the data in BigQuery to TFRecords, running a batch inference on Vertex AI Prediction, and writing the results to BigQuery requires more effort to build the inference pipeline than option A. Dataflow is a service for creating and running data processing pipelines, such as ETL (extract, transform, load) or batch processing5. Vertex AI Prediction is a service for deploying and serving ML models for online or batch prediction. However, this option also requires writing code to create the Dataflow pipeline, convert the data to TFRecords, run the batch inference, and write the results to BigQuery.
• Option D is incorrect because loading the TensorFlow SavedModel in a Dataflow pipeline, using the BigQuery I/O connector with a custom function to perform the inference within the pipeline, and writing the results to BigQuery requires more effort to build the inference pipeline than option A. The BigQuery I/O connector is a way to read and write data from BigQuery within a Dataflow pipeline. However, this option also requires writing code to load the TensorFlow SavedModel, create the custom function for inference, and write the results to BigQuery.

References:

• Importing models into BigQuery ML
• Using imported models for prediction
• TensorFlow BigQuery reader
• BigQuery API
• Dataflow overview
• [Vertex AI Prediction overview]
• [Batch prediction with Dataflow]
• [BigQuery I/O connector]
• [Using TensorFlow models in Dataflow]

• Option A is incorrect because distributing the dataset with tf.distribute.Strategy.experimental_distribute_dataset is not the most effective way to decrease the training time. This method allows you to distribute your dataset across multiple devices or machines, by creating a tf.data.Dataset instance that can be iterated over in parallel1. However, this option may not improve the training time significantly, as it does not change the amount of data or computation that each device or machine has to process. Moreover, this option may introduce additional overhead or complexity, as it requires you to handle the data sharding, replication, and synchronization across the devices or machines1.
• Option B is incorrect because creating a custom training loop is not the easiest way to decrease the training time. A custom training loop is a way to implement your own logic for training your model, by using low-level TensorFlow APIs, such as tf.GradientTape, tf.Variable, or tf.function2. A custom training loop may give you more flexibility and control over the training process, but it also requires more effort and expertise, as you have to write and debug the code for each step of the training loop, such as computing the gradients, applying the optimizer, or updating the metrics2. Moreover, a custom training loop may not improve the training time significantly, as it does not change the amount of data or computation that each device or machine has to process.
• Option C is incorrect because using a TPU with tf.distribute.TPUStrategy is not a valid way to decrease the training time. A TPU (Tensor Processing Unit) is a custom hardware accelerator designed for high-performance ML workloads3. A tf.distribute.TPUStrategy is a distribution strategy that allows you to distribute your training across multiple TPUs, by creating a tf.distribute.TPUStrategy instance that can be used with high-level TensorFlow APIs, such as Keras4. However, this option is not feasible, as Vertex AI Training does not support TPUs as accelerators for custom training jobs5. Moreover, this option may require significant code changes, as TPUs have different requirements and limitations than GPUs.
• Option D is correct because increasing the batch size is the best way to decrease the training time. The batch size is a hyperparameter that determines how many samples of data are processed in each iteration of the training loop. Increasing the batch size may reduce the training time, as it reduces the number of iterations needed to train the model, and it allows each device or machine to process more data in parallel. Increasing the batch size is also easy to implement, as it only requires changing a single hyperparameter. However, increasing the batch size may also affect the convergence and the accuracy of the model, so it is important to find the optimal batch size that balances the trade-off between the training time and the model performance.

References:

• tf.distribute.Strategy.experimental_distribute_dataset
• Custom training loop
• TPU overview
• tf.distribute.TPUStrategy
• Vertex AI Training accelerators
• [TPU programming model]
• [Batch size and learning rate]
• [Keras overview]
• [tf.distribute.MirroredStrategy]
• [Vertex AI Training overview]
• [TensorFlow overview]

Vertex AI Feature Store provides two options for online serving: Bigtable and optimized online serving. Both options support autoscaling, which means that the number of online serving nodes can automatically adjust to the traffic demand. By enabling autoscaling, you can improve the online serving performance and reduce the feature retrieval latency during the daily batch ingestion. Autoscaling also helps you optimize the cost and resource utilization of your featurestore. References:

• Online serving | Vertex AI | Google Cloud
• New Vertex AI Feature Store: BigQuery-Powered, GenAI-Ready | Google Cloud Blog

The best option for choosing a model that prioritizes detection while ensuring that more than 50% of the maintenance jobs triggered by the model address an imminent machine failure is to choose the model with the highest recall where precision is greater than 0.5. This option has the following advantages:

• It maximizes the recall, which is the proportion of actual failures that are correctly predicted by the model. Recall is also known as sensitivity or true positive rate (TPR), and it is calculated as:

mathrmRecall=fracmathrmTPmathrmTP+mathrmFN

where TP is the number of true positives (actual failures that are predicted as failures) and FN is the number of false negatives (actual failures that are predicted as non-failures). By maximizing the recall, the model can reduce the number of false negatives, which are the most costly and undesirable outcomes for the predictive maintenance use case, as they represent missed failures that can lead to machine breakdown and downtime.

• It constrains the precision, which is the proportion of predicted failures that are actual failures. Precision is also known as positive predictive value (PPV), and it is calculated as:

mathrmPrecision=fracmathrmTPmathrmTP+mathrmFP

where FP is the number of false positives (actual non-failures that are predicted as failures). By constraining the precision to be greater than 0.5, the model can ensure that more than 50% of the maintenance jobs triggered by the model address an imminent machine failure, which can avoid unnecessary or wasteful maintenance costs.

The other options are less optimal for the following reasons:

• Option A: Choosing the model with the highest area under the receiver operating characteristic curve (AUC ROC) and precision greater than 0.5 may not prioritize detection, as the AUC ROC does not directly measure the recall. The AUC ROC is a summary metric that evaluates the overall performance of a binary classifier across all possible thresholds. The ROC curve plots the TPR (recall) against the false positive rate (FPR), which is the proportion of actual non-failures that are incorrectly predicted by the model. The AUC ROC is the area under the ROC curve, and it ranges from 0 to 1, where 1 represents a perfect classifier. However, choosing the model with the highest AUC ROC may not maximize the recall, as the AUC ROC is influenced by both the TPR and the FPR, and it does not account for the precision or the specificity (the proportion of actual non-failures that are correctly predicted by the model).
• Option B: Choosing the model with the lowest root mean squared error (RMSE) and recall greater than 0.5 may not prioritize detection, as the RMSE is not a suitable metric for binary classification. The RMSE is a regression metric that measures the average magnitude of the error between the predicted and the actual values. The RMSE is calculated as:

mathrmRMSE=sqrtfrac1nsumi=1n​(yi​−hatyi​)2

where yi​ is the actual value, hatyi​ is the predicted value, and n is the number of observations. However, choosing the model with the lowest RMSE may not optimize the detection of failures, as the RMSE is sensitive to outliers and does not account for the class imbalance or the cost of misclassification.

• Option D: Choosing the model with the highest precision where recall is greater than 0.5 may not prioritize detection, as the precision may not be the most important metric for the predictive maintenance use case. The precision measures the accuracy of the positive predictions, but it does not reflect the sensitivity or the coverage of the model. By choosing the model with the highest precision, the model may sacrifice the recall, which is the proportion of actual failures that are correctly predicted by the model. This may increase the number of false negatives, which are the most costly and undesirable outcomes for the predictive maintenance use case, as they represent missed failures that can lead to machine breakdown and downtime.

References:

• Evaluation Metrics (Classifiers) - Stanford University
• Evaluation of binary classifiers - Wikipedia
• Predictive Maintenance: The greatest benefits and smart use cases

 A predict request is a way to send data to a trained model and get predictions in return. A predict request can be written in different formats, such as JSON, protobuf, or gRPC, depending on the service and the platform that are used to host and serve the model. A predict request usually contains the following information:

• The signature name: This is the name of the signature that defines the inputs and outputs of the model. A signature is a way to specify the expected format, type, and shape of the data that the model can accept and produce. A signature can be specified when exporting or saving the model, or it can be automatically inferred by the service or the platform. A model can have multiple signatures, but only one can be used for each predict request.
• The instances: This is the data that is sent to the model for prediction. The instances can be a single instance or a batch of instances, depending on the size and shape of the data. The instances should match the input specification of the signature, such as the number, name, and type of the input tensors.

For the use case of training a text classification model, the correct way to write the predict request is D. data = json.dumps({“signature_name”: “serving_default”, “instances”: [[‘a’, ‘b’], [‘c’, ‘d’], [‘e’, ‘f’]]})

This option involves writing the predict request in JSON format, which is a common and convenient format for sending and receiving data over the web. JSON stands for JavaScript Object Notation, and it is a way to represent data as a collection of name-value pairs or an ordered list of values. JSON can be easily converted to and from Python objects using the json module.

This option also involves using the signature name “serving_default”, which is the default signature name that is assigned to the model when it is saved or exported without specifying a custom signature name. The serving_default signature defines the input and output tensors of the model based on the SignatureDef that is shown in the image. According to the SignatureDef, the model expects an input tensor called “text” that has a shape of (-1, 2) and a type of DT_STRING, and produces an output tensor called “softmax” that has a shape of (-1, 2) and a type of DT_FLOAT. The -1 in the shape indicates that the dimension can vary depending on the number of instances, and the 2 indicates that the dimension is fixed at 2. The DT_STRING and DT_FLOAT indicate that the data type is string and float, respectively.

This option also involves sending a batch of three instances to the model for prediction. Each instance is a list of two strings, such as [‘a’, ‘b’], [‘c’, ‘d’], or [‘e’, ‘f’]. These instances match the input specification of the signature, as they have a shape of (3, 2) and a type of string. The model will process these instances and produce a batch of three predictions, each with a softmax output that has a shape of (1, 2) and a type of float. The softmax output is a probability distribution over the two possible classes that the model can predict, such as positive or negative sentiment.

Therefore, writing the predict request as data = json.dumps({“signature_name”: “serving_default”, “instances”: [[‘a’, ‘b’], [‘c’, ‘d’], [‘e’, ‘f’]]}) is the correct and valid way to send data to the text classification model and get predictions in return.

References:

• [json — JSON encoder and decoder]

 The best option for automating the model retraining workflow is to use GitHub Actions and Cloud Build. GitHub Actions is a service that can create and run workflows for continuous integration and continuous delivery (CI/CD) on GitHub. GitHub Actions can run tests, build and deploy code, and trigger other actions based on events such as code changes, pull requests, or manual triggers. Cloud Build is a service that can create and run scalable and reliable pipelines to build, test, and deploy software on Google Cloud. Cloud Build can build custom Docker images, push the images to Artifact Registry, and launch the pipeline in Vertex AI Pipelines. Vertex AI Pipelines is a service that can orchestrate machine learning (ML) workflows using Vertex AI. Vertex AI Pipelines can run preprocessing and training steps on custom Docker images, and evaluate, deploy, and monitor the ML model. By using GitHub Actions and Cloud Build, users can leverage the power and flexibility of Google Cloud to automate the model retraining workflow, while minimizing the steps required to build the workflow.

The other options are not as good as option D, for the following reasons:

• Option A: Triggering a Cloud Build workflow to run tests, build custom Docker images, push the images to Artifact Registry, and launch the pipeline in Vertex AI Pipelines would require more configuration and maintenance than using GitHub Actions and Cloud Build. Cloud Build is a service that can create and run pipelines to build, test, and deploy software on Google Cloud, but it is not designed to integrate with GitHub or other source code repositories. To trigger a Cloud Build workflow from GitHub, users would need to set up a webhook, a Cloud Pub/Sub topic, and a Cloud Function1. Moreover, Cloud Build does not support manual triggers, which limits the flexibility of the workflow2.
• Option B: Triggering GitHub Actions to run the tests, launching a job on Cloud Run to build custom Docker images, pushing the images to Artifact Registry, and launching the pipeline in Vertex AI Pipelines would require more steps and resources than using GitHub Actions and Cloud Build. Cloud Run is a service that can run stateless containers on a fully managed environment or on Anthos. Cloud Run can build custom Docker images, but it is not optimized for this task. Users would need to write a Dockerfile, a cloudbuild.yaml file, and a Cloud Run service configuration file, and use the gcloud command-line tool to build and deploy the image3. Moreover, Cloud Run is designed for serving HTTP requests, not for running ML pipelines, which can have different performance and scalability requirements.
• Option C: Triggering GitHub Actions to run the tests, building custom Docker images, pushing the images to Artifact Registry, and launching the pipeline in Vertex AI Pipelines would require more skills and tools than using GitHub Actions and Cloud Build. GitHub Actions can run tests and build code, but it is not specialized for building Docker images. Users would need to install and configure Docker on the GitHub Actions runner, write a Dockerfile, and use the docker command-line tool to build and push the image. Moreover, GitHub Actions has limitations on the disk space, memory, and CPU of the runner, which can affect the speed and reliability of the image building process.

References:

• Building CI/CD for Vertex AI pipelines: The first solution
• Cloud Build
• GitHub Actions
• Vertex AI Pipelines
• Triggering builds from GitHub
• Triggering builds manually
• Building containers
• Cloud Run
• [Building and testing Docker images with GitHub Actions]
• [Usage limits, billing, and administration]

 The performance of an image classification model can be measured by various metrics, such as accuracy, precision, recall, F1-score, and mean average precision (mAP). These metrics can be calculated based on the confusion matrix, which compares the predicted labels and the true labels of the images1

One of the best ways to monitor the performance of multiple versions of an image classification model on AI Platform is to compare the mean average precision across the models using the Continuous Evaluation feature. Mean average precision is a metric that summarizes the precision and recall of a model across different confidence thresholds and classes. Mean average precision is especially useful for multi-class and multi-label image classification problems, where the model has to assign one or more labels to each image from a set of possible labels. Mean average precision can range from 0 to 1, where a higher value indicates a better performance2

Continuous Evaluation is a feature of AI Platform that allows you to automatically evaluate the performance of your deployed models using online prediction requests and responses. Continuous Evaluation can help you monitor the quality and consistency of your models over time, and detect any issues or anomalies that may affect the model performance. Continuous Evaluation can also provide various evaluation metrics and visualizations, such as accuracy, precision, recall, F1-score, ROC curve, and confusion matrix, for different types of models, such as classification, regression, and object detection3

To compare the mean average precision across the models using the Continuous Evaluation feature, you need to do the following steps:

• Enable the online prediction logging for each model version that you want to evaluate. This will allow AI Platform to collect the prediction requests and responses from your models and store them in BigQuery4
• Create an evaluation job for each model version that you want to evaluate. This will allow AI Platform to compare the predicted labels and the true labels of the images, and calculate the evaluation metrics, such as mean average precision. You need to specify the BigQuery table that contains the prediction logs, the data schema, the label column, and the evaluation interval.
• View the evaluation results for each model version on the AI Platform Models page in the Google Cloud console. You can see the mean average precision and other metrics for each model version over time, and compare them using charts and tables. You can also filter the results by different classes and confidence thresholds.

The other options are not as effective or feasible. Comparing the loss performance for each model on a held-out dataset or on the validation data is not a good idea, as the loss function may not reflect the actual performance of the model on the online prediction data, and may vary depending on the choice of the loss function and the optimization algorithm. Comparing the receiver operating characteristic (ROC) curve for each model using the What-If Tool is not possible, as the What-If Tool does not support image data or multi-class classification problems.

References: 1: Confusion matrix 2: Mean average precision 3: Continuous Evaluation overview 4: Configure online prediction logging : [Create an evaluation job] : [View evaluation results] : [What-If Tool overview]

A feature cross is a synthetic feature that is obtained by combining two or more existing features, usually by taking their product or concatenation. A feature cross can help to capture the nonlinear and interaction effects between the original features, and improve the predictive performance of the model. A feature cross can be applied to different types of features, such as numeric, categorical, or geospatial features1.

For the use case of building an ML model to predict car sales in different cities around the world, the best option is to use one feature obtained as an element-wise product between binned latitude, binned longitude, and one-hot encoded car type. This option involves creating a feature cross that combines three individual features: binned latitude, binned longitude, and one-hot encoded car type. Binning is a technique that transforms a continuous numeric feature into a discrete categorical feature by dividing its range into equal intervals, or bins. One-hot encoding is a technique that transforms a categorical feature into a binary vector, where each element corresponds to a possible category, and has a value of 1 if the feature belongs to that category, and 0 otherwise. By applying binning and one-hot encoding to the latitude, longitude, and car type features, the feature cross can capture the city-specific relationships between car type and number of sales, as each combination of bins and car types can represent a different city and its preference for a certain car type. For example, the feature cross can learn that a city with a latitude bin of [40, 50], a longitude bin of [-80, -70], and a car type of SUV has a higher number of sales than a city with a latitude bin of [-10, 0], a longitude bin of [10, 20], and a car type of sedan. Therefore, using one feature obtained as an element-wise product between binned latitude, binned longitude, and one-hot encoded car type is the best option for this use case.

References:

• Feature Crosses | Machine Learning Crash Course

 Vertex AI Experiments is a managed service that allows you to track, compare, and manage experiments with Vertex AI. You can use Vertex AI Experiments to record the parameters, metrics, and artifacts of each pipeline run, and compare them in a graphical interface. Vertex AI TensorBoard is a tool that lets you visualize the metrics of your models, such as accuracy, loss, and learning curves. By logging metrics to Vertex ML Metadata and using Vertex AI Experiments and TensorBoard, you can easily collaborate with your team and find the best model configuration for your problem. References: Vertex AI Pipelines: Metrics visualization and run comparison using the KFP SDK, Track, compare, manage experiments with Vertex AI Experiments, Vertex AI Pipelines

The best option to minimize storage and computational overhead is to use the TRANSFORM clause in the CREATE MODEL statement in the SQL query to calculate the required statistics. The TRANSFORM clause allows you to specify feature preprocessing logic that applies to both training and prediction. The preprocessing logic is executed in the same query as the model creation, which avoids the need to create and store intermediate tables. The TRANSFORM clause also supports quantile bucketization and MinMax scaling, which are the preprocessing steps required for this scenario. Option A is incorrect because creating a component in the Vertex AI Pipelines DAG to calculate the required statistics may increase the computational overhead, as the component needs to run separately from the model creation. Moreover, the component needs to pass the statistics to subsequent components, which may increase the storage overhead. Option B is incorrect because preprocessing and staging the data in BigQuery prior to feeding it to the model may also increase the storage and computational overhead, as you need to create and maintain additional tables for the preprocessed data. Moreover, you need to ensure that the preprocessing logic is consistent for both training and inference. Option C is incorrect because creating SQL queries to calculate and store the required statistics in separate BigQuery tables may also increase the storage and computational overhead, as you need to create and maintain additional tables for the statistics. Moreover, you need to ensure that the statistics are updated regularly to reflect the new data. References:

• BigQuery ML documentation
• Using the TRANSFORM clause
• Feature preprocessing with BigQuery ML

BigQuery is a service that allows you to store and query large amounts of data in a scalable and cost-effective way. You can use BigQuery to ingest the Avro files from the Cloud Storage bucket and perform analytics on the structured data. Avro is a binary file format that can store complex data types and schemas. You can use the bq load command or the BigQuery API to load the Avro files into a BigQuery table. You can then use SQL queries to analyze the data and generate insights. Dataflow is a service that allows you to create and run scalable and portable data processing pipelines on Google Cloud. You can use Dataflow to create the features for your ML models, such as transforming, aggregating, and encoding the data. You can use the Apache Beam SDK to write your Dataflow pipeline code in Python or Java. You can also use the built-in transforms or custom transforms to apply the feature engineering logic to your data. Vertex AI Feature Store is a service that allows you to store and manage your ML features on Google Cloud. You can use Vertex AI Feature Store to host the features that your ML models use for online prediction. Online prediction is a type of prediction that provides low-latency responses to individual or small batches of input data. You can use the Vertex AI Feature Store API to write the features from your Dataflow pipeline to a feature store entity type. You can then use the Vertex AI Feature Store online serving API to read the features from the feature store and pass them to your ML models for online prediction. By using BigQuery, Dataflow, and Vertex AI Feature Store, you can configure a pipeline that performs analytics, creates features, and hosts the features that your ML models use for online prediction. References:

• BigQuery documentation
• Dataflow documentation
• Vertex AI Feature Store documentation
• Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate

 The best option for continuing experimenting and iterating on your pipeline to improve model performance, using Cloud Build for CI/CD, and deploying new pipelines into production quickly and easily, is to set up a CI/CD pipeline that builds and tests your source code and then deploys built artifacts into a pre-production environment. After a successful pipeline run in the pre-production environment, deploy the pipeline to production. This option allows you to leverage the power and simplicity of Cloud Build to automate, monitor, and manage your pipeline development and deployment workflow. Cloud Build is a service that can create and run continuous integration and continuous delivery (CI/CD) pipelines on Google Cloud. Cloud Build can build your source code, run unit tests, and deploy built artifacts to various Google Cloud services, such as Vertex AI Pipelines, Vertex AI Endpoints, and Artifact Registry. A CI/CD pipeline is a workflow that can automate the process of building, testing, and deploying software. A CI/CD pipeline can help you improve the quality and reliability of your software, accelerate the development and delivery cycle, and reduce the manual effort and errors. A pre-production environment is an environment that can simulate the production environment, but is isolated from the real users and data. A pre-production environment can help you test and validate your software before deploying it to production, and catch any bugs or issues that may affect the user experience or the system performance. By setting up a CI/CD pipeline that builds and tests your source code and then deploys built artifacts into a pre-production environment, you can ensure that your pipeline code is consistent and error-free, and that your pipeline artifacts are compatible and functional. After a successful pipeline run in the pre-production environment, you can deploy the pipeline to production, which is the environment where your software is accessible and usable by the real users and data. By deploying the pipeline to production after a successful pipeline run in the pre-production environment, you can minimize the chance that the new pipeline implementations will break in production, and ensure that your software meets the user expectations and requirements1.

The other options are not as good as option C, for the following reasons:

• Option A: Setting up a CI/CD pipeline that builds and tests your source code, and if the tests are successful, using the Google Cloud console to upload the built container to Artifact Registry and upload the compiled pipeline to Vertex AI Pipelines would not allow you to deploy new pipelines into production quickly and easily, and could increase the manual effort and errors. The Google Cloud console is a web-based user interface that can help you access and manage various Google Cloud services, such as Artifact Registry and Vertex AI Pipelines. Artifact Registry is a service that can store and manage your container images and other artifacts on Google Cloud. Artifact Registry can help you upload and organize your container images, and track the image versions and metadata. Vertex AI Pipelines is a service that can orchestrate machine learning workflows using Vertex AI. Vertex AI Pipelines can run preprocessing and training steps on custom Docker images, and evaluate, deploy, and monitor the machine learning model. However, setting up a CI/CD pipeline that builds and tests your source code, and if the tests are successful, using the Google Cloud console to upload the built container to Artifact Registry and upload the compiled pipeline to Vertex AI Pipelines would not allow you to deploy new pipelines into production quickly and easily, and could increase the manual effort and errors. You would need to write code, create and run the CI/CD pipeline, use the Google Cloud console to upload the built container to Artifact Registry, and use the Google Cloud console to upload the compiled pipeline to Vertex AI Pipelines. Moreover, this option would not use a pre-production environment to test and validate your pipeline before deploying it to production, which could increase the chance that the new pipeline implementations will break in production1.
• Option B: Setting up a CI/CD pipeline that builds your source code and then deploys built artifacts into a pre-production environment, running unit tests in the pre-production environment, and if the tests are successful, deploying the pipeline to production would not allow you to test and validate your pipeline before deploying it to production, and could cause errors or poor performance. A unit test is a type of test that can verify the functionality and correctness of a small and isolated unit of code, such as a function or a class. A unit test can help you debug and improve your code quality, and catch any bugs or issues that may affect the code logic or output. However, setting up a CI/CD pipeline that builds your source code and then deploys built artifacts into a pre-production environment, running unit tests in the pre-production environment, and if the tests are successful, deploying the pipeline to production would not allow you to test and validate your pipeline before deploying it to production, and could cause errors or poor performance. You would need to write code, create and run the CI/CD pipeline, deploy the built artifacts to the pre-production environment, run the unit tests in the pre-production environment, and deploy the pipeline to production. Moreover, this option would not run the pipeline in the pre-production environment, which could prevent you from testing and validating the pipeline functionality and compatibility, and catching any bugs or issues that may affect the pipeline workflow or output1.
• Option D: Setting up a CI/CD pipeline that builds and tests your source code and then deploys built artifacts into a pre-production environment, after a successful pipeline run in the pre-production environment, rebuilding the source code, and deploying the artifacts to production would not allow you to deploy new pipelines into production quickly and easily, and could increase the complexity and cost of the pipeline development and deployment. Rebuilding the source code is a process that can recompile and repackage the source code into executable artifacts, such as container images and pipeline files. Rebuilding the source code can help you incorporate any changes or updates that may have occurred in the source code, and ensure that the artifacts are consistent and up-to-date. However, setting up a CI/CD pipeline that builds and tests your source code and then deploys built artifacts into a pre-production environment, after a successful pipeline run in the pre-production environment, rebuilding the source code, and deploying the artifacts to production would not allow you to deploy new pipelines into production quickly and easily, and could increase the complexity and cost of the pipeline development and deployment. You would need to write code, create and run the CI/CD pipeline, deploy the built artifacts to the pre-production environment, run the pipeline in the pre-production environment, rebuild the source code, and deploy the artifacts to production. Moreover, this option would increase the pipeline development and deployment time, as rebuilding the source code can be a time-consuming and resource-intensive process1.

References:

• Preparing for Google Cloud Certification: Machine Learning Engineer, Course 3: Production ML Systems, Week 3: MLOps
• Google Cloud Professional Machine Learning Engineer Exam Guide, Section 3: Scaling ML models in production, 3.2 Automating ML workflows
• Official Google Cloud Certified Professional Machine Learning Engineer Study Guide, Chapter 6: Production ML Systems, Section 6.4: Automating ML Workflows
• Cloud Build
• Vertex AI Pipelines
• Artifact Registry
• Pre-production environment

 A TPU VM is a virtual machine that has direct access to a Cloud TPU device. TPU VMs provide a simpler and more flexible way to use Cloud TPUs, as they eliminate the need for a separate host VM and network setup. TPU VMs also support interactive debugging tools such as TensorFlow Debugger (tfdbg) and Python Debugger (pdb), which can help researchers develop and troubleshoot complex models. A v3-8 TPU VM has 8 TPU cores, which can provide high performance and scalability for training large models. SSHing into the TPU VM allows the user to run and debug the TensorFlow code directly on the TPU device, without any network overhead or data transfer issues. References:

• 1: TPU VMs Overview
• 2: TPU VMs Quickstart
• 3: Debugging TensorFlow Models on Cloud TPUs

The best option for dealing with the missing categorical variable in the test set is to apply one-hot encoding on the categorical variables in the test data. This option has the following advantages:

• It ensures the consistency and compatibility of the data format for the ML model, as the one-hot encoding transforms the categorical variables into binary vectors that can be easily processed by the model. By applying one-hot encoding on the categorical variables in the test data, you can match the number and order of the features in the test data with the training data, and avoid any errors or discrepancies in the model prediction.
• It preserves the information and relevance of the data for the ML model, as the one-hot encoding creates a separate feature for each possible value of the categorical variable, and assigns a value of 1 to the feature corresponding to the actual value of the variable, and 0 to the rest. By applying one-hot encoding on the categorical variables in the test data, you can retain the original meaning and importance of the categorical variable, and avoid any loss or distortion of the data.

The other options are less optimal for the following reasons:

• Option A: Randomly redistributing the data, with 70% for the training set and 30% for the test set, introduces additional complexity and risk. This option requires reshuffling and splitting the data again, which can be tedious and time-consuming. Moreover, this option may not guarantee that the missing categorical variable will be present in the test set, as it depends on the randomness of the data distribution. Furthermore, this option may affect the quality and validity of the ML model, as it may change the data characteristics and patterns that the model has learned from the original training set.
• Option B: Using sparse representation in the test set introduces additional overhead and inefficiency. This option requires converting the categorical variables in the test set into sparse vectors, which are vectors that have mostly zero values and only store the indices and values of the non-zero elements. However, using sparse representation in the test set may not be compatible with the ML model, as the model expects the input data to have the same format and dimensionality as the training data, which uses one-hot encoding. Moreover, using sparse representation in the test set may not be efficient or scalable, as it requires additional computation and memory to store and process the sparse vectors.
• Option D: Collecting more data representing all categories introduces additional cost and delay. This option requires obtaining and labeling more data that contains the missing categorical variable, which can be expensive and time-consuming. Moreover, this option may not be feasible or necessary, as the missing categorical variable may not be available or relevant for the test data, depending on the data source or the business problem.

According to the official exam guide1, one of the skills assessed in the exam is to “explain the predictions of a trained model”. Vertex AI provides feature attributions using Shapley Values, a cooperative game theory algorithm that assigns credit to each feature in a model for a particular outcome2. Feature attributions can help you understand how the model calculates the predictions and debug or optimize the model accordingly. You can use AutoML for Tabular Data to generate and query local feature attributions3. The other options are not relevant or optimal for this scenario. References:

• Professional ML Engineer Exam Guide
• Feature attributions for classification and regression
• AutoML for Tabular Data
• Google Professional Machine Learning Certification Exam 2023
• Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

 Entity extraction is a natural language processing (NLP) task that involves identifying and extracting specific types of information from text, such as names, dates, locations, etc. Entity extraction can help you analyze a corpus of recipes and extract each ingredient and cookware mentioned in them. Vertex AI is a unified platform for building and managing machine learning solutions on Google Cloud. It provides a service for AutoML entity extraction, which allows you to create and train custom entity extraction models without writing any code. You can use Vertex AI to create a text dataset for entity extraction, and label your data with two entities: “ingredient” and “cookware”. You need to label at least 200 examples of each entity type to train an AutoML entity extraction model. You can also use a holdout dataset to evaluate the performance of your model, such as precision, recall, and F1-score. This solution can help you build a machine learning model to scan a corpus of recipes and extract each ingredient and cookware mentioned in them, and use the results to help users with meal planning. References:

• AutoML Entity Extraction | Vertex AI
• Preparing data for AutoML Entity Extraction | Vertex AI

 According to the official exam guide1, one of the skills assessed in the exam is to “explain the predictions of a trained model”. TensorFlow2 is an open source framework for developing and deploying machine learning and deep learning models. TensorFlow supports various model explainability methods, such as Integrated Gradients3, which is a technique that assigns an importance score to each input feature by approximating the integral of the gradients along the path from a baseline input to the actual input. Integrated Gradients can help explain the output of a deep learning-based model by highlighting the most influential features in the input images. Therefore, option C is the best way to build the model for the given use case. The other options are not relevant or optimal for this scenario. References:

• Professional ML Engineer Exam Guide
• TensorFlow
• Integrated Gradients
• Google Professional Machine Learning Certification Exam 2023
• Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

• Option A is not the best answer because it requires using Apache Spark and Dataproc, which may incur additional cost and complexity for running and managing the cluster. It also requires saving the preprocessed data as CSV files in a Cloud Storage bucket, which may increase the storage cost and the data transfer latency.
• Option B is not the best answer because it requires loading the data into a pandas DataFrame, which may not be scalable or efficient for large datasets. It also requires training the model directly on the DataFrame, which may not leverage the distributed computing capabilities of BigQuery.
• Option C is the best answer because it allows performing preprocessing in BigQuery by using SQL, which is a cost-effective and efficient way to manipulate large datasets. It also allows using the BigQueryClient in TensorFlow to read the data directly from BigQuery, which is a convenient and fast way to access the data for training the model1.
• Option D is not the best answer because it requires using Apache Beam and Dataflow, which may incur additional cost and complexity for running and managing the pipeline. It also requires saving the preprocessed data as CSV files in a Cloud Storage bucket, which may increase the storage cost and the data transfer latency.

References:

• 1: Read data from BigQuery | TensorFlow I/O

• A fairness metric is a way to measure how well a machine learning model treats different groups of customers, such as by sex or age. A common fairness metric is accuracy, which is the proportion of correct predictions among all predictions. Accuracy across the sensitive features means calculating the accuracy for each group separately, and then comparing them. For example, if the model has 90% accuracy for male customers and 80% accuracy for female customers, there is a 10% accuracy gap that indicates potential bias against female customers.
• To investigate and mitigate potential bias, it is important to define a fairness metric and evaluate it on a test set. A test set is a subset of the data that is not used for training the model, but only for evaluating its performance. By joining the test set predictions with the sensitive features, you can calculate the fairness metric and see if it meets your requirements. For example, you may require that the accuracy gap between any two groups is less than 5%. If the fairness metric does not meet your requirements, you may need to adjust the model or the data to reduce bias.
• Option A is not the best answer because excluding the sensitive features and any meaningfully correlated features may not eliminate bias. For example, if salary is correlated with sex, and salary is also a predictive feature for the target variable, excluding both features may reduce the model accuracy and still leave some residual bias. Moreover, excluding features based on correlation may not capture the complex interactions and dependencies among the features that may affect bias.
• Option B is not the best answer because using the global attribution values for each feature of the model may not reflect the individual-level impact of the features on the predictions. Global attribution values are calculated by averaging the attribution values across all the data points, and they indicate how important each feature is for the overall model performance. However, they do not show how each feature affects each customer’s prediction, which may vary depending on the values of the other features. For example, sex may have a low global attribution value, but it may have a high impact on some customers’ predictions, especially if it interacts with other features such as salary or age.
• Option C is not the best answer because discarding the model and training the model again without a feature based on a single customer’s attribution value may not be a robust or scalable way to mitigate bias. Attribution values are calculated by measuring how much each feature contributes to the prediction for a given data point, and they indicate how sensitive the prediction is to the feature value. However, they do not show how the feature affects the overall fairness metric or the model accuracy. For example, sex may have a high attribution value for a customer, but it may not affect the accuracy gap between the groups. Moreover, discarding and retraining the model based on a single customer’s attribution value may not be feasible if there are many customers with high attribution values for different features.

AI Platform Training is a service that allows you to run your machine learning experiments on Google Cloud using various features, model architectures, and hyperparameters. You can use AI Platform Training to scale up your experiments, leverage distributed training, and access specialized hardware such as GPUs and TPUs1. Cloud Monitoring is a service that collects and analyzes metrics, logs, and traces from Google Cloud, AWS, and other sources. You can use Cloud Monitoring to create dashboards, alerts, and reports based on your data2. The Monitoring API is an interface that allows you to programmatically access and manipulate your monitoring data3.

By using AI Platform Training and Cloud Monitoring, you can track and report your experiments while minimizing manual effort. You can write the accuracy metrics from your experiments to Cloud Monitoring using the AI Platform Training Python package4. You can then query the results using the Monitoring API and compare the performance of different experiments. You can also visualize the metrics in the Cloud Console or create custom dashboards and alerts5. Therefore, using AI Platform Training and Cloud Monitoring is the best option for this use case.

References:

• AI Platform Training documentation
• Cloud Monitoring documentation
• Monitoring API overview
• Using Cloud Monitoring with AI Platform Training
• Viewing evaluation metrics

 According to the official exam guide1, one of the skills assessed in the exam is to “track the lineage of pipeline artifacts”. Vertex ML Metadata2 is a service that allows you to store, query, and visualize metadata associated with your ML workflows, such as datasets, models, metrics, and executions. Vertex ML Metadata helps you track the provenance and lineage of your ML artifacts and understand the relationships between them. Vertex AI Experiments3 is a service that allows you to track and compare the results of your model training runs. Vertex AI Experiments automatically logs metadata such as hyperparameters, metrics, and artifacts for each training run. You can use Vertex AI Experiments to train your custom model using TensorFlow, PyTorch, XGBoost, or scikit-learn. Vertex AI TensorBoard4 is a service that allows you to visualize and monitor your ML experiments using TensorBoard, an open source tool for ML visualization. Vertex AI TensorBoard helps you track the model’s training parameters and the metrics per training epoch, and compare the performance of each version of the model. Therefore, option C is the best way to determine which Vertex AI services to include in the workflow for the given use case. The other options are not relevant or optimal for this scenario. References:

• Professional ML Engineer Exam Guide
• Vertex ML Metadata
• Vertex AI Experiments
• Vertex AI TensorBoard
• Google Professional Machine Learning Certification Exam 2023
• Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

AI Platform Training is a service that allows you to run your machine learning training jobs on Google Cloud using various features, model architectures, and hyperparameters. You can use AI Platform Training to scale up your training jobs, leverage distributed training, and access specialized hardware such as GPUs and TPUs1. AI Platform Training supports several pre-built containers that provide different ML frameworks and dependencies, such as TensorFlow, PyTorch, scikit-learn, and XGBoost2. However, if the ML framework and related dependencies that you need are not supported by the pre-built containers, you can build your own custom containers and use them to run your training jobs on AI Platform Training3.

Custom containers are Docker images that you create to run your training application. By using custom containers, you can specify and pre-install all the dependencies needed for your application, and have full control over the code, serving, and deployment of your model4. Custom containers also enable you to run distributed training jobs on AI Platform Training, which can help you train large-scale and complex models faster and more efficiently5. Distributed training is a technique that splits the training data and computation across multiple machines, and coordinates them to update the model parameters. AI Platform Training supports two types of distributed training: parameter server and collective all-reduce. The parameter server architecture consists of a set of workers that perform the computation, and a set of servers that store and update the model parameters. The collective all-reduce architecture consists of a set of workers that perform the computation and synchronize the model parameters among themselves. Both architectures also have a scheduler that coordinates the workers and servers.

For the use case of training a custom neural network that uses critical dependencies specific to your organization’s framework, the best option is to build your custom containers to run distributed training jobs on AI Platform Training. This option allows you to use the ML framework and dependencies of your choice, and train your model on multiple machines without having to manage the infrastructure. Since your ML framework of choice uses the scheduler, workers, and servers distribution structure, you can use the parameter server architecture to run your distributed training job on AI Platform Training. You can specify the number and type of machines, the custom container image, and the training application arguments when you submit your training job. Therefore, building your custom containers to run distributed training jobs on AI Platform Training is the best option for this use case.

References:

• AI Platform Training documentation
• Pre-built containers for training
• Custom containers for training
• Custom containers overview | Vertex AI | Google Cloud
• Distributed training overview
• [Types of distributed training]
• [Distributed training architectures]
• [Using custom containers for training with the parameter server architecture]

A Dataflow streaming pipeline is a cost-effective way to process large volumes of real-time data from sensors. The RunInference API is a Dataflow transform that allows you to run online predictions on your streaming data using your ML models. By using the RunInference API, you can avoid the latency and cost of using a separate prediction service. The automatic model refresh feature enables you to update your models in the pipeline without redeploying the pipeline. This way, you can ensure that your models are always up-to-date and accurate. By deploying a Dataflow streaming pipeline with the RunInference API and using automatic model refresh, you can achieve sub-millisecond predictions, 24/7 availability, and low operational overhead for your ML models. References:

• Dataflow documentation
• RunInference API documentation
• Automatic model refresh documentation
• Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate

According to the web search results, TensorFlow Extended (TFX) is a platform for building end-to-end machine learning pipelines using TensorFlow1. TFX provides a set of components that can be orchestrated using either the TFX SDK or Kubeflow Pipelines. TFX components can handle different aspects of the pipeline, such as data ingestion, data validation, data transformation, model training, model evaluation, model serving, and more. TFX components can also leverage other Google Cloud services, such as Dataflow2 and Vertex AI3. Dataflow is a fully managed service for running Apache Beam pipelines on Google Cloud. Dataflow handles the provisioning and management of the compute resources, as well as the optimization and execution of the pipelines. Vertex AI is a unified platform for machine learning development and deployment. Vertex AI offers various services and tools for building, managing, and serving machine learning models. Therefore, option D is the best way to create a low maintenance, automated workflow for the given use case, as it allows you to use the TFX SDK to define and execute your pipeline components, and use Dataflow and Vertex AI services to scale and optimize your pipeline. The other options are not relevant or optimal for this scenario. References:

• TensorFlow Extended
• Dataflow
• Vertex AI
• Google Professional Machine Learning Certification Exam 2023
• Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

 The best options for adjusting the training parameters in AutoML to improve model performance are to decrease the score threshold and add more positive examples to the training set. These options can help increase the detection rate of fraudulent transactions, which is the priority for this use case. The score threshold is a parameter that determines the minimum probability score that a prediction must have to be classified as positive. Decreasing the score threshold can increase the recall of the model, which is the proportion of actual positive cases that are correctly identified. Increasing the recall can help reduce the number of false negatives, which are fraudulent transactions that are missed by the model. However, decreasing the score threshold can also decrease the precision of the model, which is the proportion of positive predictions that are actually correct. Decreasing the precision can increase the number of false positives, which are legitimate transactions that are flagged as fraudulent by the model. Therefore, there is a trade-off between recall and precision, and the optimal score threshold depends on the business objective and the cost of errors1. Adding more positive examples to the training set can help balance the data distribution and improve the model performance. Positive examples are the instances that belong to the target class, which in this case are fraudulent transactions. Negative examples are the instances that belong to the other class, which in this case are legitimate transactions. Fraudulent transactions are usually rare and imbalanced compared to legitimate transactions, which can cause the model to be biased towards the majority class and fail to learn the characteristics of the minority class. Adding more positive examples can help the model learn more features and patterns of the fraudulent transactions, and increase the detection rate2.

The other options are not as good as options B and C, for the following reasons:

• Option A: Increasing the score threshold would decrease the detection rate of fraudulent transactions, which is the opposite of the desired outcome. Increasing the score threshold would decrease the recall of the model, which is the proportion of actual positive cases that are correctly identified. Decreasing the recall would increase the number of false negatives, which are fraudulent transactions that are missed by the model. Increasing the score threshold would increase the precision of the model, which is the proportion of positive predictions that are actually correct. Increasing the precision would decrease the number of false positives, which are legitimate transactions that are flagged as fraudulent by the model. However, in this use case, the cost of false negatives is much higher than the cost of false positives, so increasing the score threshold is not a good option1.
• Option D: Adding more negative examples to the training set would not improve the model performance, and could worsen the data imbalance. Negative examples are the instances that belong to the other class, which in this case are legitimate transactions. Legitimate transactions are usually abundant and dominant compared to fraudulent transactions, which can cause the model to be biased towards the majority class and fail to learn the characteristics of the minority class. Adding more negative examples would exacerbate this problem, and decrease the detection rate of the fraudulent transactions2.
• Option E: Reducing the maximum number of node hours for training would not improve the model performance, and could limit the model optimization. Node hours are the units of computation that are used to train an AutoML model. The maximum number of node hours is a parameter that determines the upper limit of node hours that can be used for training. Reducing the maximum number of node hours would reduce the training time and cost, but also the model quality and accuracy. Reducing the maximum number of node hours would limit the number of iterations, trials, and evaluations that the model can perform, and prevent the model from finding the optimal hyperparameters and architecture3.

References:

• Preparing for Google Cloud Certification: Machine Learning Engineer, Course 5: Responsible AI, Week 4: Evaluation
• Google Cloud Professional Machine Learning Engineer Exam Guide, Section 2: Developing high-quality ML models, 2.2 Handling imbalanced data
• Official Google Cloud Certified Professional Machine Learning Engineer Study Guide, Chapter 4: Low-code ML Solutions, Section 4.3: AutoML
• Understanding the score threshold slider
• Handling imbalanced data sets in machine learning
• AutoML Vision pricing

 A data pipeline is a set of steps or processes that move data from one or more sources to one or more destinations, usually for the purpose of analysis, transformation, or storage. A data pipeline can be designed using various components, such as data sources, data processing tools, data storage systems, and data analytics tools1

To design a data pipeline for analyzing customer sentiments in each call, one should consider the following requirements and constraints:

• The call center receives over one million calls daily, and data is stored in Cloud Storage. This implies that the data is large, unstructured, and distributed, and requires a scalable and efficient data processing tool that can handle various types of data formats, such as audio, text, or image.
• The data collected must not leave the region in which the call originated, and no Personally Identifiable Information (Pll) can be stored or analyzed. This implies that the data is sensitive and subject to data privacy and compliance regulations, and requires a secure and reliable data storage system that can enforce data encryption, access control, and regional policies.
• The data science team has a third-party tool for visualization and access which requires a SQL ANSI-2011 compliant interface. This implies that the data analytics tool is external and independent of the data pipeline, and requires a standard and compatible data interface that can support SQL queries and operations.

One of the best options for selecting components for data processing and for analytics is to use Dataflow for data processing and BigQuery for analytics. Dataflow is a fully managed service for executing Apache Beam pipelines for data processing, such as batch or stream processing, extract-transform-load (ETL), or data integration. BigQuery is a serverless, scalable, and cost-effective data warehouse that allows you to run fast and complex queries on large-scale data23

Using Dataflow and BigQuery has several advantages for this use case:

• Dataflow can process large and unstructured data from Cloud Storage in a parallel and distributed manner, and apply various transformations, such as converting audio to text, extracting sentiment scores, or anonymizing PII. Dataflow can also handle both batch and stream processing, which can enable real-time or near-real-time analysis of the call data.
• BigQuery can store and analyze the processed data from Dataflow in a secure and reliable way, and enforce data encryption, access control, and regional policies. BigQuery can also support SQL ANSI-2011 compliant interface, which can enable the data science team to use their third-party tool for visualization and access. BigQuery can also integrate with various Google Cloud services and tools, such as AI Platform, Data Studio, or Looker.
• Dataflow and BigQuery can work seamlessly together, as they are both part of the Google Cloud ecosystem, and support various data formats, such as CSV, JSON, Avro, or Parquet. Dataflow and BigQuery can also leverage the benefits of Google Cloud infrastructure, such as scalability, performance, and cost-effectiveness.

The other options are not as suitable or feasible. Using Pub/Sub for data processing and Datastore for analytics is not ideal, as Pub/Sub is mainly designed for event-driven and asynchronous messaging, not data processing, and Datastore is mainly designed for low-latency and high-throughput key-value operations, not analytics. Using Cloud Function for data processing and Cloud SQL for analytics is not optimal, as Cloud Function has limitations on the memory, CPU, and execution time, and does not support complex data processing, and Cloud SQL is a relational database service that may not scale well for large-scale data. Using Cloud Composer for data processing and Cloud SQL for analytics is not relevant, as Cloud Composer is mainly designed for orchestrating complex workflows across multiple systems, not data processing, and Cloud SQL is a relational database service that may not scale well for large-scale data.

References: 1: Data pipeline 2: Dataflow overview 3: BigQuery overview : [Dataflow documentation] : [BigQuery documentation]

 BigQuery is a service that allows you to store and query large amounts of data in a scalable and cost-effective way. You can use BigQuery to build a model that predicts customer lifetime value over the next three years, by using the create model statement. The create model statement is a SQL command that allows you to create and train an ML model using your data in BigQuery. You can use the create model statement to create an AutoML model, which is a type of model that automatically selects the best features and architecture for your data. By using an AutoML model, you can use the simplest approach to build the model, without writing any code or performing any feature engineering. You can also use the ml.evaluate and ml.predict statements to validate the results of your model. The ml.evaluate statement is a SQL command that allows you to evaluate the performance and quality of your model using various metrics. The ml.predict statement is a SQL command that allows you to make predictions using your model and new data. You can also use the BigQuery console to access visualization tools, such as charts and graphs, to explore and analyze your data and model results. By using the BigQuery console, the create model statement, and the ml.evaluate and ml.predict statements, you can build and validate a model that predicts customer lifetime value over the next three years, and have access to visualization tools. References:

• BigQuery documentation
• create model statement documentation
• ml.evaluate statement documentation
• ml.predict statement documentation
• Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate

The best option for deploying a custom tabular ML model to production for online predictions, and monitoring the feature distribution over time with minimal effort, using a model that was provided by the bank’s vendor, the training data is not available due to its sensitivity, and the model is packaged as a Vertex AI Model serving container which accepts a string as input for each prediction instance, is to upload the model to Vertex AI Model Registry and deploy the model to a Vertex AI endpoint, create a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective, and provide an instance schema. This option allows you to leverage the power and simplicity of Vertex AI to serve and monitor your model with minimal code and configuration. Vertex AI is a unified platform for building and deploying machine learning solutions on Google Cloud. Vertex AI can deploy a trained model to an online prediction endpoint, which can provide low-latency predictions for individual instances. Vertex AI can also provide various tools and services for data analysis, model development, model deployment, model monitoring, and model governance. A Vertex AI Model Registry is a resource that can store and manage your models on Vertex AI. A Vertex AI Model Registry can help you organize and track your models, and access various model information, such as model name, model description, and model labels. A Vertex AI Model serving container is a resource that can run your custom model code on Vertex AI. A Vertex AI Model serving container can help you package your model code and dependencies into a container image, and deploy the container image to an online prediction endpoint. A Vertex AI Model serving container can accept various input formats, such as JSON, CSV, or TFRecord. A string input format is a type of input format that accepts a string as input for each prediction instance. A string input format can help you encode your feature values into a single string, and separate them by commas. By uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, you can serve your model for online predictions with minimal code and configuration. You can use the Vertex AI API or the gcloud command-line tool to upload the model to Vertex AI Model Registry, and provide the model name, model description, and model labels. You can also use the Vertex AI API or the gcloud command-line tool to deploy the model to a Vertex AI endpoint, and provide the endpoint name, endpoint description, endpoint labels, and endpoint resources. A Vertex AI Model Monitoring job is a resource that can monitor the performance and quality of your deployed models on Vertex AI. A Vertex AI Model Monitoring job can help you detect and diagnose issues with your models, such as data drift, prediction drift, training/serving skew, or model staleness. Feature drift is a type of model monitoring metric that measures the difference between the distributions of the features used to train the model and the features used to serve the model over time. Feature drift can indicate that the online data is changing over time, and the model performance is degrading. By creating a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective, and providing an instance schema, you can monitor the feature distribution over time with minimal effort. You can use the Vertex AI API or the gcloud command-line tool to create a Vertex AI Model Monitoring job, and provide the monitoring objective, the monitoring frequency, the alerting threshold, and the notification channel. You can also provide an instance schema, which is a JSON file that describes the features and their types in the prediction input data. An instance schema can help Vertex AI Model Monitoring parse and analyze the string input format, and calculate the feature distributions and distance scores1.

The other options are not as good as option A, for the following reasons:

• Option B: Uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature skew detection as the monitoring objective, and providing an instance schema would not help you monitor the changes in the online data over time, and could cause errors or poor performance. Feature skew is a type of model monitoring metric that measures the difference between the distributions of the features used to train the model and the features used to serve the model at a given point in time. Feature skew can indicate that the model is not trained on the representative data, or that the data is changing over time. By creating a Vertex AI Model Monitoring job with feature skew detection as the monitoring objective, and providing an instance schema, you can monitor the feature distribution at a given point in time with minimal effort. However, uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature skew detection as the monitoring objective, and providing an instance schema would not help you monitor the changes in the online data over time, and could cause errors or poor performance. You would need to use the Vertex AI API or the gcloud command-line tool to upload the model to Vertex AI Model Registry, deploy the model to a Vertex AI endpoint, create a Vertex AI Model Monitoring job, and provide an instance schema. Moreover, this option would not monitor the feature drift, which is a more direct and relevant metric for measuring the changes in the online data over time, and the model performance and quality1.
• Option C: Refactoring the serving container to accept key-value pairs as input format, uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective would require more skills and steps than uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective, and providing an instance schema. A key-value pair input format is a type of input format that accepts a key-value pair as input for each prediction instance. A key-value pair input format can help you specify the feature names and values in a JSON object, and separate them by colons. By refactoring the serving container to accept key-value pairs as input format, uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective, you can serve and monitor your model with minimal code and configuration. You can write code to refactor the serving container to accept key-value pairs as input format, and use the Vertex AI API or the gcloud command-line tool to upload the model to Vertex AI Model Registry, deploy the model to a Vertex AI endpoint, and create a Vertex AI Model Monitoring job. However, refactoring the serving container to accept key-value pairs as input format, uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective would require more skills and steps than uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective, and providing an instance schema. You would need to write code, refactor the serving container, upload the model to Vertex AI Model Registry, deploy the model to a Vertex AI endpoint, and create a Vertex AI Model Monitoring job. Moreover, this option would not use the instance schema, which is a JSON file that can help Vertex AI Model Monitoring parse and analyze the string input format, and calculate the feature distributions and distance scores1.
• Option D: Refactoring the serving container to accept key-value pairs as input format, uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature skew detection as the monitoring objective would require more skills and steps than uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective, and providing an instance schema, and would not help you monitor the changes in the online data over time, and could cause errors or poor performance. Feature skew is a type of model monitoring metric that measures the difference between the distributions of the features used to train the model and the features used to serve the model at a given point in time. Feature skew can indicate that the model is not trained on the representative data, or that the data is changing over time. By creating a Vertex AI Model Monitoring job with feature skew detection as the monitoring objective, you can monitor the feature distribution at a given point in time with minimal effort. However, refactoring the serving container to accept key-value pairs as input format, uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature skew detection as the monitoring objective would require more skills and steps than uploading the model to Vertex AI Model Registry and deploying the model to a Vertex AI endpoint, creating a Vertex AI Model Monitoring job with feature drift detection as the monitoring objective, and providing an instance schema, and would not help you monitor the changes in the online data over time, and could cause errors or poor performance. You would need to write code, refactor the serving container, upload the model to Vertex AI Model Registry, deploy the model to a Vertex AI endpoint, and create a Vertex AI Model Monitoring job. Moreover, this option would not monitor the feature drift, which is a more direct and relevant metric for measuring the changes in the online data over time, and the model performance and quality1.

References:

• Using Model Monitoring | Vertex AI | Google Cloud

 A cloud-based backend system is a system that runs on a cloud platform and provides services or resources to other applications or users. A cloud-based backend system can be used to submit training jobs, which are tasks that involve training a machine learning model on a given dataset using a specific framework and configuration1

However, a cloud-based backend system can also have some drawbacks, such as:

• High maintenance: A cloud-based backend system may require a lot of administration and management, such as provisioning, scaling, monitoring, and troubleshooting the cloud resources and services. This can be time-consuming and costly, and may distract from the core business objectives2
• Low flexibility: A cloud-based backend system may not support all the frameworks and libraries that the data scientists need to use for their training jobs. This can limit the choices and capabilities of the data scientists, and affect the quality and performance of their models3
• Poor integration: A cloud-based backend system may not integrate well with other cloud services or tools that the data scientists need to use for their machine learning workflows, such as data processing, model deployment, or model monitoring. This can create compatibility and interoperability issues, and reduce the efficiency and productivity of the data scientists.

Therefore, it may be better to use a managed service instead of a cloud-based backend system to submit training jobs. A managed service is a service that is provided and operated by a third-party provider, and offers various benefits, such as:

• Low maintenance: A managed service handles the administration and management of the cloud resources and services, and abstracts away the complexity and details of the underlying infrastructure. This can save time and money, and allow the data scientists to focus on their core tasks2
• High flexibility: A managed service can support multiple frameworks and libraries that the data scientists need to use for their training jobs, and allow them to customize and configure their training environments and parameters. This can enhance the choices and capabilities of the data scientists, and improve the quality and performance of their models3
• Easy integration: A managed service can integrate seamlessly with other cloud services or tools that the data scientists need to use for their machine learning workflows, and provide a unified and consistent interface and experience. This can solve the compatibility and interoperability issues, and increase the efficiency and productivity of the data scientists.

One of the best options for using a managed service to submit training jobs is to use the AI Platform custom containers feature to receive training jobs using any framework. AI Platform is a Google Cloud service that provides a platform for building, deploying, and managing machine learning models. AI Platform supports various machine learning frameworks, such as TensorFlow, PyTorch, scikit-learn, and XGBoost, and provides various features, such as hyperparameter tuning, distributed training, online prediction, and model monitoring.

The AI Platform custom containers feature allows the data scientists to use any framework or library that they want for their training jobs, and package their training application and dependencies as a Docker container image. The data scientists can then submit their training jobs to AI Platform, and specify the container image and the training parameters. AI Platform will run the training jobs on the cloud infrastructure, and handle the scaling, logging, and monitoring of the training jobs. The data scientists can also use the AI Platform features to optimize, deploy, and manage their models.

The other options are not as suitable or feasible. Configuring Kubeflow to run on Google Kubernetes Engine and receive training jobs through TFJob is not ideal, as Kubeflow is mainly designed for TensorFlow-based training jobs, and does not support other frameworks or libraries. Creating a library of VM images on Compute Engine and publishing these images on a centralized repository is not optimal, as Compute Engine is a low-level service that requires a lot of administration and management, and does not provide the features and integrations of AI Platform. Setting up Slurm workload manager to receive jobs that can be scheduled to run on your cloud infrastructure is not relevant, as Slurm is a tool for managing and scheduling jobs on a cluster of nodes, and does not provide a managed service for training jobs.

References: 1: Cloud computing 2: Managed services 3: Machine learning frameworks : [Machine learning workflow] : [AI Platform overview] : [Custom containers for training]

A ResourceExhaustedError: out of memory (OOM) when allocating tensor is an error that occurs when the GPU runs out of memory while trying to allocate memory for a tensor. A tensor is a multi-dimensional array of numbers that represents the data or the parameters of a machine learning model. The size and shape of a tensor depend on various factors, such as the input data, the model architecture, the batch size, and the optimization algorithm1.

For the use case of training a computer vision model that predicts the type of government ID present in a given image using a GPU-powered virtual machine on Compute Engine, the best option to resolve the error is to reduce the batch size. The batch size is a parameter that determines how many input examples are processed at a time by the model. A larger batch size can improve the model’s accuracy and stability, but it also requires more memory and computation. A smaller batch size can reduce the memory and computation requirements, but it may also affect the model’s performance and convergence2.

By reducing the batch size, the GPU can allocate less memory for each tensor, and avoid running out of memory. Reducing the batch size can also speed up the training process, as the GPU can process more batches in parallel. However, reducing the batch size too much may also have some drawbacks, such as increasing the noise and variance of the gradient updates, and slowing down the convergence of the model. Therefore, the optimal batch size should be chosen based on the trade-off between memory, computation, and performance3.

The other options are not as effective as option B, because they are not directly related to the memory allocation of the GPU. Option A, changing the optimizer, may affect the speed and quality of the optimization process, but it may not reduce the memory usage of the model. Option C, changing the learning rate, may affect the convergence and stability of the model, but it may not reduce the memory usage of the model. Option D, reducing the image shape, may reduce the size of the input tensor, but it may also reduce the quality and resolution of the image, and affect the model’s accuracy. Therefore, option B, reducing the batch size, is the best answer for this question.

References:

• ResourceExhaustedError: OOM when allocating tensor with shape - Stack Overflow
• How does batch size affect model performance and training time? - Stack Overflow
• How to choose an optimal batch size for training a neural network? - Stack Overflow

The best option for resolving the training-serving skew alert is to update the model monitoring job to use the more recent training data that was used to retrain the model. This option can help align the baseline distribution of the model monitoring job with the current distribution of the production data, and eliminate the false positive alerts. Model Monitoring is a service that can track and compare the results of multiple machine learning runs. Model Monitoring can monitor the model’s prediction input data for feature skew and drift. Training-serving skew occurs when the feature data distribution in production deviates from the feature data distribution used to train the model. If the original training data is available, you can enable skew detection to monitor your models for training-serving skew. Model Monitoring uses TensorFlow Data Validation (TFDV) to calculate the distributions and distance scores for each feature, and compares them with a baseline distribution. The baseline distribution is the statistical distribution of the feature’s values in the training data. If the distance score for a feature exceeds an alerting threshold that you set, Model Monitoring sends you an email alert. However, if you retrain the model with more recent training data, and deploy it back to the Vertex AI endpoint, the baseline distribution of the model monitoring job may become outdated and inconsistent with the current distribution of the production data. This can cause the model monitoring job to generate false positive alerts, even if the model performance is not deteriorated. To avoid this problem, you need to update the model monitoring job to use the more recent training data that was used to retrain the model. This can help the model monitoring job to recalculate the baseline distribution and the distance scores, and compare them with the current distribution of the production data. This can also help the model monitoring job to detect any true positive alerts, such as a sudden change in the production data that causes the model performance to degrade1.

The other options are not as good as option B, for the following reasons:

• Option A: Updating the model monitoring job to use a lower sampling rate would not resolve the training-serving skew alert, and could reduce the accuracy and reliability of the model monitoring job. The sampling rate is a parameter that determines the percentage of prediction requests that are logged and analyzed by the model monitoring job. Using a lower sampling rate can reduce the storage and computation costs of the model monitoring job, but also the quality and validity of the data. Using a lower sampling rate can introduce sampling bias and noise into the data, and make the model monitoring job miss some important features or patterns of the data. Moreover, using a lower sampling rate would not address the root cause of the training-serving skew alert, which is the mismatch between the baseline distribution and the current distribution of the production data2.
• Option C: Temporarily disabling the alert, and enabling the alert again after a sufficient amount of new production traffic has passed through the Vertex AI endpoint, would not resolve the training-serving skew alert, and could expose the model to potential risks and errors. Disabling the alert would stop the model monitoring job from sending email notifications when the distance score for a feature exceeds the alerting threshold, but it would not stop the model monitoring job from calculating and comparing the distributions and distance scores. Therefore, disabling the alert would not address the root cause of the training-serving skew alert, which is the mismatch between the baseline distribution and the current distribution of the production data. Moreover, disabling the alert would prevent the model monitoring job from detecting any true positive alerts, such as a sudden change in the production data that causes the model performance to degrade. This can expose the model to potential risks and errors, and affect the user satisfaction and trust1.
• Option D: Temporarily disabling the alert until the model can be retrained again on newer training data, and retraining the model again after a sufficient amount of new production traffic has passed through the Vertex AI endpoint, would not resolve the training-serving skew alert, and could cause unnecessary costs and efforts. Disabling the alert would stop the model monitoring job from sending email notifications when the distance score for a feature exceeds the alerting threshold, but it would not stop the model monitoring job from calculating and comparing the distributions and distance scores. Therefore, disabling the alert would not address the root cause of the training-serving skew alert, which is the mismatch between the baseline distribution and the current distribution of the production data. Moreover, disabling the alert would prevent the model monitoring job from detecting any true positive alerts, such as a sudden change in the production data that causes the model performance to degrade. This can expose the model to potential risks and errors, and affect the user satisfaction and trust. Retraining the model again on newer training data would create a new model version, but it would not update the model monitoring job to use the newer training data as the baseline distribution. Therefore, retraining the model again on newer training data would not resolve the training-serving skew alert, and could cause unnecessary costs and efforts1.

References:

• Preparing for Google Cloud Certification: Machine Learning Engineer, Course 3: Production ML Systems, Week 4: Evaluation
• Google Cloud Professional Machine Learning Engineer Exam Guide, Section 3: Scaling ML models in production, 3.3 Monitoring ML models in production
• Official Google Cloud Certified Professional Machine Learning Engineer Study Guide, Chapter 6: Production ML Systems, Section 6.3: Monitoring ML Models
• Using Model Monitoring
• Understanding the score threshold slider
• Sampling rate

 The best option for splitting the data between the training, validation, and test sets, using a managed tabular dataset in Vertex AI that contains sales data from three different stores, is to use Vertex AI default data split. This option allows you to leverage the power and simplicity of Vertex AI to automatically and randomly split your data into the three sets by percentage. Vertex AI is a unified platform for building and deploying machine learning solutions on Google Cloud. Vertex AI can support various types of models, such as linear regression, logistic regression, k-means clustering, matrix factorization, and deep neural networks. Vertex AI can also provide various tools and services for data analysis, model development, model deployment, model monitoring, and model governance. A default data split is a data split method that is provided by Vertex AI, and does not require any user input or configuration. A default data split can help you split your data into the training, validation, and test sets by using a random sampling method, and assign a fixed percentage of the data to each set. A default data split can help you simplify the data split process, and works well in most cases. A training set is a subset of the data that is used to train the model, and adjust the model parameters. A training set can help you learn the relationship between the input features and the target variable, and optimize the model performance. A validation set is a subset of the data that is used to validate the model, and tune the model hyperparameters. A validation set can help you evaluate the model performance on unseen data, and avoid overfitting or underfitting. A test set is a subset of the data that is used to test the model, and provide the final evaluation metrics. A test set can help you assess the model performance on new data, and measure the generalization ability of the model. By using Vertex AI default data split, you can split your data into the training, validation, and test sets by using a random sampling method, and assign the following percentages of the data to each set1:

The other options are not as good as option B, for the following reasons:

• Option A: Using Vertex AI manual split, using the store name feature to assign one store for each set would not allow you to split your data into representative and balanced sets, and could cause errors or poor performance. A manual split is a data split method that allows you to control how your data is split into sets, by using the ml_use label or the data filter expression. A manual split can help you customize the data split logic, and handle complex or non-standard data formats. A store name feature is a feature that indicates the name of the store where the sales data was collected. A store name feature can help you identify the source of the data, and group the data by store. However, using Vertex AI manual split, using the store name feature to assign one store for each set would not allow you to split your data into representative and balanced sets, and could cause errors or poor performance. You would need to write code, create and configure the ml_use label or the data filter expression, and assign one store for each set. Moreover, this option would not ensure that the data in each set has the same distribution and characteristics as the data in the whole dataset, which could prevent you from learning the general pattern of the data, and cause bias or variance in the model2.
• Option C: Using Vertex AI chronological split and specifying the sales timestamp feature as the time variable would not allow you to split your data into representative and balanced sets, and could cause errors or poor performance. A chronological split is a data split method that allows you to split your data into sets based on the order of the data. A chronological split can help you preserve the temporal dependency and sequence of the data, and avoid data leakage. A sales timestamp feature is a feature that indicates the date and time when the sales data was collected. A sales timestamp feature can help you track the changes and trends of the data over time, and capture the seasonality and cyclicality of the data. However, using Vertex AI chronological split and specifying the sales timestamp feature as the time variable would not allow you to split your data into representative and balanced sets, and could cause errors or poor performance. You would need to write code, create and configure the time variable, and split the data by the order of the time variable. Moreover, this option would not ensure that the data in each set has the same distribution and characteristics as the data in the whole dataset, which could prevent you from learning the general pattern of the data, and cause bias or variance in the model3.
• Option D: Using Vertex AI random split, assigning 70% of the rows to the training set, 10% to the validation set, and 20% to the test set would not allow you to use the default data split method that is provided by Vertex AI, and could increase the complexity and cost of the data split process. A random split is a data split method that allows you to split your data into sets by using a random sampling method, and assign a custom percentage of the data to each set. A random split can help you split your data into representative and balanced sets, and avoid data leakage. However, using Vertex AI random split, assigning 70% of the rows to the training set, 10% to the validation set, and 20% to the test set would not allow you to use the default data split method that is provided by Vertex AI, and could increase the complexity and cost of the data split process. You would need to write code, create and configure the random split method, and assign the custom percentages to each set. Moreover, this option would not use the default data split method that is provided by Vertex AI, which can simplify the data split process, and works well in most cases1.

References:

• About data splits for AutoML models | Vertex AI | Google Cloud
• Manual split for unstructured data
• Mathematical split

TFRecords are a binary file format that can store large amounts of data efficiently. By converting the images to TFRecords and storing them in a Cloud Storage bucket, you can reduce the data size and improve the data transfer speed. You can then read the TFRecords by using the tf.data.TFRecordDataset function, which creates a dataset of tensors from the TFRecord files. This way, you can read images at scale during training while minimizing data I/O bottlenecks. References:

• TFRecord documentation
• tf.data.TFRecordDataset documentation
• Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate

 The tf.data dataset is a TensorFlow API that provides a way to create and manipulate data pipelines for machine learning. The tf.data dataset allows you to apply various transformations to the data, such as reading, shuffling, batching, prefetching, and interleaving. These transformations can affect the performance and efficiency of the model training process1

One of the common performance issues in model training is input-bound, which means that the model is waiting for the input data to be ready and is not fully utilizing the computational resources. Input-bound can be caused by slow data loading, insufficient parallelism, or large data size. Input-bound can be detected by using the Cloud TPU profiler plugin, which is a tool that helps you analyze the performance of your model on Cloud TPUs. The Cloud TPU profiler plugin can show you the percentage of time that the TPU cores are idle, which indicates input-bound2

To reduce the input-bound bottleneck and speed up the model training process, you can make some modifications to the tf.data dataset. Two of the modifications that can help are:

• Use the interleave option for reading data. The interleave option allows you to read data from multiple files in parallel and interleave their records. This can improve the data loading speed and reduce the idle time of the TPU cores. The interleave option can be applied by using the tf.data.Dataset.interleave method, which takes a function that returns a dataset for each input element, and a number of parallel calls3
• Set the prefetch option equal to the training batch size. The prefetch option allows you to prefetch the next batch of data while the current batch is being processed by the model. This can reduce the latency between batches and improve the throughput of the model training. The prefetch option can be applied by using the tf.data.Dataset.prefetch method, which takes a buffer size argument. The buffer size should be equal to the training batch size, which is the number of examples per batch4

The other options are not effective or counterproductive. Reducing the value of the repeat parameter will reduce the number of epochs, which is the number of times the model sees the entire dataset. This can affect the model’s accuracy and convergence. Increasing the buffer size for the shuffle option will increase the randomness of the data, but also increase the memory usage and the data loading time. Decreasing the batch size argument in your transformation will reduce the number of examples per batch, which can affect the model’s stability and performance.

References: 1: tf.data: Build TensorFlow input pipelines 2: Cloud TPU Tools in TensorBoard 3: tf.data.Dataset.interleave 4: tf.data.Dataset.prefetch : [Better performance with the tf.data API]

• Option A is incorrect because creating a new model that uses the raw data and is available in real time would require retraining the model and deploying it again, which is not efficient or scalable.
• Option B is incorrect because using a Dataflow job to transform the incoming data would introduce unnecessary latency and complexity for online prediction, which requires fast and simple processing.
• Option C is incorrect because using Cloud Spanner to stream and query the incoming data would incur high costs and overhead for online prediction, which does not need a relational database.
• Option D is correct because using a Cloud Function to preprocess the data and submit a prediction request to Al Platform is a simple and scalable solution for online prediction, which leverages the serverless and event-driven features of Cloud Functions.

 Vertex AI Workbench is an integrated development environment (IDE) that allows you to create and run Jupyter notebooks on Google Cloud. Vertex AI Pipelines is a service that allows you to create and manage machine learning workflows using Vertex AI components. To submit a Vertex AI Pipeline job from a Vertex AI Workbench instance, you need to have the appropriate permissions to access the Vertex AI resources. The Identity and Access Management (IAM) Vertex AI User role is a predefined role that grants the minimum permissions required to use Vertex AI services, such as creating and deploying models, endpoints, and pipelines. By assigning the Vertex AI User role to the Vertex AI Workbench instance, you can ensure that the instance has sufficient permissions to submit a Vertex AI Pipeline job. You can assign the role to the instance by using the Cloud Console, the gcloud command-line tool, or the Cloud IAM API. References: The answer can be verified from official Google Cloud documentation and resources related to Vertex AI Workbench, Vertex AI Pipelines, and IAM.

• Vertex AI Workbench | Google Cloud
• Vertex AI Pipelines | Google Cloud
• Vertex AI roles | Google Cloud
• Granting, changing, and revoking access to resources | Google Cloud

The best option for adding explanations to your model code with minimal effort and providing explanations that are as accurate as possible is to upload the custom model to Vertex AI Model Registry and configure feature-based attribution by using sampled Shapley with input baselines. This option allows you to leverage the power and simplicity of Vertex Explainable AI to generate feature attributions for each prediction, and understand how each feature contributes to the model output. Vertex Explainable AI is a service that can help you understand and interpret predictions made by your machine learning models, natively integrated with a number of Google’s products and services. Vertex Explainable AI can provide feature-based and example-based explanations to provide better understanding of model decision making. Feature-based explanations are explanations that show how much each feature in the input influenced the prediction. Feature-based explanations can help you debug and improve model performance, build confidence in the predictions, and understand when and why things go wrong. Vertex Explainable AI supports various feature attribution methods, such as sampled Shapley, integrated gradients, and XRAI. Sampled Shapley is a feature attribution method that is based on the Shapley value, which is a concept from game theory that measures how much each player in a cooperative game contributes to the total payoff. Sampled Shapley approximates the Shapley value for each feature by sampling different subsets of features, and computing the marginal contribution of each feature to the prediction. Sampled Shapley can provide accurate and consistent feature attributions, but it can also be computationally expensive. To reduce the computation cost, you can use input baselines, which are reference inputs that are used to compare with the actual inputs. Input baselines can help you define the starting point or the default state of the features, and calculate the feature attributions relative to the input baselines. By uploading the custom model to Vertex AI Model Registry and configuring feature-based attribution by using sampled Shapley with input baselines, you can add explanations to your model code with minimal effort and provide explanations that are as accurate as possible1.

The other options are not as good as option C, for the following reasons:

• Option A: Creating an AutoML tabular model by using the BigQuery data with integrated Vertex Explainable AI would require more skills and steps than uploading the custom model to Vertex AI Model Registry and configuring feature-based attribution by using sampled Shapley with input baselines. AutoML tabular is a service that can automatically build and train machine learning models for structured or tabular data. AutoML tabular can use BigQuery as the data source, and provide feature-based explanations by using integrated gradients as the feature attribution method. However, creating an AutoML tabular model by using the BigQuery data with integrated Vertex Explainable AI would require more skills and steps than uploading the custom model to Vertex AI Model Registry and configuring feature-based attribution by using sampled Shapley with input baselines. You would need to create a new AutoML tabular model, import the BigQuery data, configure the model settings, train and evaluate the model, and deploy the model. Moreover, this option would not use your existing custom model, which is already performing well, but create a new model, which may not have the same performance or behavior as your custom model2.
• Option B: Creating a BigQuery ML deep neural network model, and using the ML.EXPLAIN_PREDICT method with the num_integral_steps parameter would not allow you to deploy the model to production, and could provide less accurate explanations than using sampled Shapley with input baselines. BigQuery ML is a service that can create and train machine learning models by using SQL queries on BigQuery. BigQuery ML can create a deep neural network model, which is a type of machine learning model that consists of multiple layers of neurons, and can learn complex patterns and relationships from the data. BigQuery ML can also provide feature-based explanations by using the ML.EXPLAIN_PREDICT method, which is a SQL function that returns the feature attributions for each prediction. The ML.EXPLAIN_PREDICT method uses integrated gradients as the feature attribution method, which is a method that calculates the average gradient of the prediction output with respect to the feature values along the path from the input baseline to the input. The num_integral_steps parameter is a parameter that determines the number of steps along the path from the input baseline to the input. However, creating a BigQuery ML deep neural network model, and using the ML.EXPLAIN_PREDICT method with the num_integral_steps parameter would not allow you to deploy the model to production, and could provide less accurate explanations than using sampled Shapley with input baselines. BigQuery ML does not support deploying the model to Vertex AI Endpoints, which is a service that can provide low-latency predictions for individual instances. BigQuery ML only supports batch prediction, which is a service that can provide high-throughput predictions for a large batch of instances. Moreover, integrated gradients can provide less accurate and consistent explanations than sampled Shapley, as integrated gradients can be sensitive to the choice of the input baseline and the num_integral_steps parameter3.
• Option D: Updating the custom serving container to include sampled Shapley-based explanations in the prediction outputs would require more skills and steps than uploading the custom model to Vertex AI Model Registry and configuring feature-based attribution by using sampled Shapley with input baselines. A custom serving container is a container image that contains the model, the dependencies, and a web server. A custom serving container can help you customize the prediction behavior of your model, and handle complex or non-standard data formats. However, updating the custom serving container to include sampled Shapley-based explanations in the prediction outputs would require more skills and steps than uploading the custom model to Vertex AI Model Registry and configuring feature-based attribution by using sampled Shapley with input baselines. You would need to write code, implement the sampled Shapley algorithm, build and test the container image, and upload and deploy the container image. Moreover, this option would not leverage the power and simplicity of Vertex Explainable AI, which can provide feature-based explanations natively integrated with Vertex AI services4.

References:

• Preparing for Google Cloud Certification: Machine Learning Engineer, Course 3: Production ML Systems, Week 4: Evaluation
• Google Cloud Professional Machine Learning Engineer Exam Guide, Section 3: Scaling ML models in production, 3.3 Monitoring ML models in production
• Official Google Cloud Certified Professional Machine Learning Engineer Study Guide, Chapter 6: Production ML Systems, Section 6.3: Monitoring ML Models
• Vertex Explainable AI
• AutoML Tables
• BigQuery ML
• Using custom containers for prediction

The best way to operationalize your training process is to use Vertex AI Pipelines, which allows you to create and run scalable, portable, and reproducible workflows for your ML models. Vertex AI Pipelines also integrates with Vertex AI Metadata, which tracks the provenance, lineage, and artifacts of your ML models. By using a Vertex AI CustomTrainingJobOp component, you can train your model using the same code as in your Jupyter notebook. By using a ModelUploadOp component, you can upload your trained model to Vertex AI Model Registry, which manages the versions and endpoints of your models. By using Cloud Scheduler and Cloud Functions, you can trigger your Vertex AI pipeline to run weekly, according to your plan. References:

• Vertex AI Pipelines documentation
• Vertex AI Metadata documentation
• Vertex AI CustomTrainingJobOp documentation
• ModelUploadOp documentation
• Cloud Scheduler documentation
• [Cloud Functions documentation]