PMLE practice questions 191–200

1. Question 191 · 364 in bank

   You recently created a new Google Cloud project. After testing that you can submit a Vertex AI Pipeline job from the Cloud Shell, you want to use a Vertex AI Workbench user-managed notebook instance to run your code from that instance. You created the instance and ran the code but this time the job fails with an insufficient permissions error. What should you do?

   • AEnsure that the Workbench instance that you created is in the same region of the Vertex AI Pipelines resources you will use.
   • BEnsure that the Vertex AI Workbench instance is on the same subnetwork of the Vertex AI Pipeline resources that you will use.
   • CEnsure that the Vertex AI Workbench instance is assigned the Identity and Access Management (IAM) Vertex AI User role.
   • DEnsure that the Vertex AI Workbench instance is assigned the Identity and Access Management (IAM) Notebooks Runner role.

   Show answer

   Answer: C

   Explanation

   I decided to change my mind to C after realizing we need the permissions aiplatform.pipelineJobs, present in vertex AI user. Not sure if the notebook runner role allows to run notebook from pipeline jobs + its specified that it only is allowed to run scheduled notebooks (no mention of scheduling here anywhere)

   Source: Adapted from community discussion — b1a8fae

   References

   • Google Cloud documentation

   Next question

2. Question 192 · 364 in bank

   You work for a semiconductor manufacturing company. You need to create a real-time application that automates the quality control process. High-definition images of each semiconductor are taken at the end of the assembly line in real time. The photos are uploaded to a Cloud Storage bucket along with tabular data that includes each semiconductor’s batch number, serial number, dimensions, and weight. You need to configure model training and serving while maximizing model accuracy. What should you do?

   • AUse Vertex AI Data Labeling Service to label the images, and tram an AutoML image classification model. Deploy the model, and configure Pub/Sub to publish a message when an image is categorized into the failing class.
   • BUse Vertex AI Data Labeling Service to label the images, and train an AutoML image classification model. Schedule a daily batch prediction job that publishes a Pub/Sub message when the job completes.
   • CConvert the images into an embedding representation. Import this data into BigQuery, and train a BigQuery ML K-means clustering model with two clusters. Deploy the model and configure Pub/Sub to publish a message when a semiconductor’s data is categorized into the failing cluster.
   • DImport the tabular data into BigQuery, use Vertex AI Data Labeling Service to label the data and train an AutoML tabular classification model. Deploy the model, and configure Pub/Sub to publish a message when a semiconductor’s data is categorized into the failing class.

   Show answer

   Answer: A

   Explanation

   Real-time Processing, uploading images to Cloud Storage triggers the AutoML image classification model for immediate processing, enabling real-time quality control decisions. Image Classification, the scenario focuses on classifying images as "passing" or "failing" quality, making image classification the appropriate approach. Pub/Sub Notifications, Pub/Sub messaging efficiently alerts downstream systems about failing classifications, allowing for prompt quality control actions.

   Source: Adapted from community discussion — omermahgoub

   Next question

3. Question 193 · 364 in bank

   You work for a rapidly growing social media company. Your team builds TensorFlow recommender models in an on-premises CPU cluster. The data contains billions of historical user events and 100,000 categorical features. You notice that as the data increases, the model training time increases. You plan to move the models to Google Cloud. You want to use the most scalable approach that also minimizes training time. What should you do?

   • ADeploy the training jobs by using TPU VMs with TPUv3 Pod slices, and use the TPUEmbeading API
   • BDeploy the training jobs in an autoscaling Google Kubernetes Engine cluster with CPUs
   • CDeploy a matrix factorization model training job by using BigQuery ML
   • DDeploy the training jobs by using Compute Engine instances with A100 GPUs, and use the tf.nn.embedding_lookup API

   Show answer

   Answer: A

   Explanation

   TPU (Tensor Processing Units) VMs are specialized hardware accelerators designed by Google specifically for machine learning tasks. TPUv3 Pod slices offer high scalability and are excellent for distributed training tasks. The TPUEmbedding API is optimized for handling large volumes of categorical features, which fits your scenario with 100,000 categorical features. This option is likely to offer the fastest training times due to specialized hardware and optimized APIs for large-scale machine learning tasks.

   Source: Adapted from community discussion — daidai75

   Next question

4. Question 194 · 364 in bank

   You are training and deploying updated versions of a regression model with tabular data by using Vertex AI Pipelines, Vertex AI Training, Vertex AI Experiments, and Vertex AI Endpoints. The model is deployed in a Vertex AI endpoint, and your users call the model by using the Vertex AI endpoint. You want to receive an email when the feature data distribution changes significantly, so you can retrigger the training pipeline and deploy an updated version of your model. What should you do?

   • AUse Vertex Al Model Monitoring. Enable prediction drift monitoring on the endpoint, and specify a notification email.
   • BIn Cloud Logging, create a logs-based alert using the logs in the Vertex Al endpoint. Configure Cloud Logging to send an email when the alert is triggered.
   • CIn Cloud Monitoring create a logs-based metric and a threshold alert for the metric. Configure Cloud Monitoring to send an email when the alert is triggered.
   • DExport the container logs of the endpoint to BigQuery. Create a Cloud Function to run a SQL query over the exported logs and send an email. Use Cloud Scheduler to trigger the Cloud Function.

   Show answer

   Answer: A

   References

   • Google Cloud documentation

   Next question

5. Question 195 · 364 in bank

   You have trained an XGBoost model that you plan to deploy on Vertex AI for online prediction. You are now uploading your model to Vertex AI Model Registry, and you need to configure the explanation method that will serve online prediction requests to be returned with minimal latency. You also want to be alerted when feature attributions of the model meaningfully change over time. What should you do?

   • A1. Specify sampled Shapley as the explanation method with a path count of 5. 2. Deploy the model to Vertex AI Endpoints. 3. Create a Model Monitoring job that uses prediction drift as the monitoring objective.
   • B1. Specify Integrated Gradients as the explanation method with a path count of 5. 2. Deploy the model to Vertex AI Endpoints. 3. Create a Model Monitoring job that uses prediction drift as the monitoring objective.
   • C1. Specify sampled Shapley as the explanation method with a path count of 50. 2. Deploy the model to Vertex AI Endpoints. 3. Create a Model Monitoring job that uses training-serving skew as the monitoring objective.
   • D1. Specify Integrated Gradients as the explanation method with a path count of 50. 2. Deploy the model to Vertex AI Endpoints. 3. Create a Model Monitoring job that uses training-serving skew as the monitoring objective.

   Show answer

   Answer: A

   Explanation

   Explanation Method: Sampled Shapley: This method provides high-fidelity feature attributions while being computationally efficient, making it ideal for low-latency online predictions. Integrated Gradients: While also accurate, it's generally more computationally intensive than sampled Shapley, potentially introducing latency. Path Count: Lower Path Count (5): Reducing path count further decreases computation time, optimizing for faster prediction responses. Monitoring Objective: Prediction Drift: This type of monitoring detects changes in feature importance over time, aligning with the goal of tracking feature attribution shifts. Training-Serving Skew: This monitors discrepancies between training and serving data distributions, which isn't directly related to feature attributions.

   Source: Adapted from community discussion — pikachu007

   References

   • Google Cloud documentation

   Next question

6. Question 196 · 364 in bank

   You work at a gaming startup that has several terabytes of structured data in Cloud Storage. This data includes gameplay time data, user metadata, and game metadata. You want to build a model that recommends new games to users that requires the least amount of coding. What should you do?

   • ALoad the data in BigQuery. Use BigQuery ML to train an Autoencoder model.
   • BLoad the data in BigQuery. Use BigQuery ML to train a matrix factorization model.
   • CRead data to a Vertex AI Workbench notebook. Use TensorFlow to train a two-tower model.
   • DRead data to a Vertex AI Workbench notebook. Use TensorFlow to train a matrix factorization model.

   Show answer

   Answer: B

   Explanation

   Using BigQuery ML for training a matrix factorization model would require less coding compared to building a custom model with TensorFlow in a Vertex AI Workbench notebook. BigQuery ML provides high-level APIs for machine learning tasks directly within the BigQuery environment, thus reducing the amount of coding needed for data preprocessing and model training. Matrix factorization is a commonly used technique for recommendation systems, making it a suitable choice for recommending new games to users based on their gameplay time data, user metadata, and game metadata.

   Source: Adapted from community discussion — guilhermebutzke

   References

   • Google developer documentation

   Next question

7. Question 197 · 364 in bank

   You work for a large bank that serves customers through an application hosted in Google Cloud that is running in the US and Singapore. You have developed a PyTorch model to classify transactions as potentially fraudulent or not. The model is a three-layer perceptron that uses both numerical and categorical features as input, and hashing happens within the model. You deployed the model to the us-central1 region on nl-highcpu-16 machines, and predictions are served in real time. The model's current median response latency is 40 ms. You want to reduce latency, especially in Singapore, where some customers are experiencing the longest delays. What should you do?

   • AAttach an NVIDIA T4 GPU to the machines being used for online inference.
   • BChange the machines being used for online inference to nl-highcpu-32.
   • CDeploy the model to Vertex AI private endpoints in the us-central1 and asia-southeast1 regions, and allow the application to choose the appropriate endpoint.
   • DCreate another Vertex AI endpoint in the asia-southeast1 region, and allow the application to choose the appropriate endpoint.

   Show answer

   Answer: C

   Explanation

   My Answer: C The bottleneck is network latency. So, A: Not Correct: might improve performance, but it's an expensive solution and may not be necessary if the bottleneck is network latency. B: Not Correct: might offer slight improvement, but the primary issue is geographical distance between users and the model. C: CORRECT: This approach leverages the geographical proximity of the endpoints to the users, reducing latency for customers in Singapore without neglecting customers in the US. Additionally, using Vertex AI private endpoints ensures secure and efficient communication between the application and the model. D: Not Correct: it's not the most efficient approach because it does not utilize the existing infrastructure in the us-central1 region, and managing multiple endpoints might introduce additional complexity.

   Source: Adapted from community discussion — guilhermebutzke

   Next question

8. Question 198 · 364 in bank

   You need to train an XGBoost model on a small dataset. Your training code requires custom dependencies. You want to minimize the startup time of your training job. How should you set up your Vertex AI custom training job?

   • AStore the data in a Cloud Storage bucket, and create a custom container with your training application. In your training application, read the data from Cloud Storage and train the model.
   • BUse the XGBoost prebuilt custom container. Create a Python source distribution that includes the data and installs the dependencies at runtime. In your training application, load the data into a pandas DataFrame and train the model.
   • CCreate a custom container that includes the data. In your training application, load the data into a pandas DataFrame and train the model.
   • DStore the data in a Cloud Storage bucket, and use the XGBoost prebuilt custom container to run your training application. Create a Python source distribution that installs the dependencies at runtime. In your training application, read the data from Cloud Storage and train the model.

   Show answer

   Answer: A

   Explanation

   Given the focus on minimizing startup time, and based on the information about XGBoost prebuilt container dependencies available here https://cloud.google.com/vertex-ai/docs/training/pre-built-containers#xgboost A: Separate Data and Custom Container is the best approach for minimizing startup time, especially for small datasets. Separating data in Cloud Storage keeps the container image lean, leading to faster download and startup compared to bundling data within the container. B. The prebuilt Container could have unnecessary components, potentially increasing the image size and impacting startup time.

   Source: Adapted from community discussion — omermahgoub

   References

   • Google Cloud documentation

   Next question

9. Question 199 · 364 in bank

   You are creating an ML pipeline for data processing, model training, and model deployment that uses different Google Cloud services. You have developed code for each individual task, and you expect a high frequency of new files. You now need to create an orchestration layer on top of these tasks. You only want this orchestration pipeline to run if new files are present in your dataset in a Cloud Storage bucket. You also want to minimize the compute node costs. What should you do?

   • ACreate a pipeline in Vertex AI Pipelines. Configure the first step to compare the contents of the bucket to the last time the pipeline was run. Use the scheduler API to run the pipeline periodically.
   • BCreate a Cloud Function that uses a Cloud Storage trigger and deploys a Cloud Composer directed acyclic graph (DAG).
   • CCreate a pipeline in Vertex AI Pipelines. Create a Cloud Function that uses a Cloud Storage trigger and deploys the pipeline.
   • DDeploy a Cloud Composer directed acyclic graph (DAG) with a GCSObjectUpdateSensor class that detects when a new file is added to the Cloud Storage bucket.

   Show answer

   Answer: C

   Explanation

   Option C appears to be the best choice for balancing the requirements of efficient orchestration, cost minimization, and ensuring the pipeline only runs when new files are present. By using a Cloud Function triggered by Cloud Storage events to deploy a Vertex AI Pipeline, you can leverage the event-driven model of Cloud Functions to minimize unnecessary runs and associated costs, while still using the powerful orchestration capabilities of Vertex AI Pipelines.

   Source: Adapted from community discussion — fitri001

   Next question

10. Question 200 · 364 in bank

    Select 2

    You are training a Resnet model on AI Platform using TPUs to visually categorize types of defects in automobile engines. You capture the training profile using the Cloud TPU profiler plugin and observe that it is highly input-bound. You want to reduce the bottleneck and speed up your model training process. Which modifications should you make to the tf.data dataset? (Choose two.)

    • AUse the interleave option for reading data.
    • BReduce the value of the repeat parameter.
    • CIncrease the buffer size for the shuttle option.
    • DSet the prefetch option equal to the training batch size.
    • EDecrease the batch size argument in your transformation.

    Show answer

    Answer: AD

    Explanation

    A. Use the interleave option for reading data. - Yes, that helps to parallelize data reading. B. Reduce the value of the repeat parameter. - No, this is only to repeat rows of the dataset. C. Increase the buffer size for the shuttle option. - No, there is only a shuttle option. D. Set the prefetch option equal to the training batch size. - Yes, this will pre-load the data. E. Decrease the batch size argument in your transformation. - No, could be even slower due to more I/Os. https://www.tensorflow.org/guide/data_performance

    Source: Adapted from community discussion — danielp14021990

    References

    • Google Cloud documentation

    Next question

Keyboard:J Kmove between questionsA–Dselect an answerRshow or hide the answer← →change page

PMLE practice, question 191. 364 questions available.