PMLE practice questions 41–50

1. Question 41 · 364 in bank

   You work for the AI team of an automobile company, and you are developing a visual defect detection model using TensorFlow and Keras. To improve your model performance, you want to incorporate some image augmentation functions such as translation, cropping, and contrast tweaking. You randomly apply these functions to each training batch. You want to optimize your data processing pipeline for run time and compute resources utilization. What should you do?

   • AEmbed the augmentation functions dynamically in the tf.Data pipeline.
   • BEmbed the augmentation functions dynamically as part of Keras generators.
   • CUse Dataflow to create all possible augmentations, and store them as TFRecords.
   • DUse Dataflow to create the augmentations dynamically per training run, and stage them as TFRecords.

   Show answer

   Answer: A

   Explanation

   A is best because, 1. It allows you to apply the augmentations on-the-fly during training, which eliminates the need for pre-processing and storing a large number of augmented images. This saves both storage space and compute resources. 2. The tf.Data pipeline is highly optimized for efficient data loading and processing, ensuring that your model training process is not bottlenecked by data preprocessing. 3. By applying augmentations randomly to each training batch, you increase the diversity of your training data, which can help your model generalize better to unseen data. Keras generators can be used for data augmentation, but tf.Data pipelines are generally more efficient and flexible for creating complex data processing pipelines.

   Source: Adapted from community discussion — tavva_prudhvi

   Next question

2. Question 42 · 364 in bank

   You work for an online publisher that delivers news articles to over 50 million readers. You have built an AI model that recommends content for the company’s weekly newsletter. A recommendation is considered successful if the article is opened within two days of the newsletter’s published date and the user remains on the page for at least one minute. All the information needed to compute the success metric is available in BigQuery and is updated hourly. The model is trained on eight weeks of data, on average its performance degrades below the acceptable baseline after five weeks, and training time is 12 hours. You want to ensure that the model’s performance is above the acceptable baseline while minimizing cost. How should you monitor the model to determine when retraining is necessary?

   • AUse Vertex AI Model Monitoring to detect skew of the input features with a sample rate of 100% and a monitoring frequency of two days.
   • BSchedule a cron job in Cloud Tasks to retrain the model every week before the newsletter is created.
   • CSchedule a weekly query in BigQuery to compute the success metric.
   • DSchedule a daily Dataflow job in Cloud Composer to compute the success metric.

   Show answer

   Answer: C

   Explanation

   Option C is the best answer. Since all the information needed to compute the success metric is available in BigQuery and is updated hourly, scheduling a weekly query in BigQuery to compute the success metric is the simplest and most cost-effective way to monitor the model's performance. By comparing the computed success metric against the acceptable baseline, you can determine when the model's performance has degraded below the threshold, and retrain the model accordingly. This approach avoids the cost of additional monitoring infrastructure and leverages existing data processing capabilities.

   Source: Adapted from community discussion — TNT87

   References

   • Google Cloud documentation

   Next question

3. Question 43 · 364 in bank

   You deployed an ML model into production a year ago. Every month, you collect all raw requests that were sent to your model prediction service during the previous month. You send a subset of these requests to a human labeling service to evaluate your model’s performance. After a year, you notice that your model's performance sometimes degrades significantly after a month, while other times it takes several months to notice any decrease in performance. The labeling service is costly, but you also need to avoid large performance degradations. You want to determine how often you should retrain your model to maintain a high level of performance while minimizing cost. What should you do?

   • ATrain an anomaly detection model on the training dataset, and run all incoming requests through this model. If an anomaly is detected, send the most recent serving data to the labeling service.
   • BIdentify temporal patterns in your model’s performance over the previous year. Based on these patterns, create a schedule for sending serving data to the labeling service for the next year.
   • CCompare the cost of the labeling service with the lost revenue due to model performance degradation over the past year. If the lost revenue is greater than the cost of the labeling service, increase the frequency of model retraining; otherwise, decrease the model retraining frequency.
   • DRun training-serving skew detection batch jobs every few days to compare the aggregate statistics of the features in the training dataset with recent serving data. If skew is detected, send the most recent serving data to the labeling service.

   Show answer

   Answer: D

   Explanation

   Option D is the best approach to determine how often to retrain the model while minimizing cost. Running training-serving skew detection batch jobs every few days to compare the aggregate statistics of the features in the training dataset with recent serving data is an effective way to detect when the model's performance has degraded. If skew is detected, the most recent serving data should be sent to the labeling service to evaluate the model's performance. This approach is more cost-effective than sending a subset of requests to the labeling service every month because it only sends data when there is a high probability that the model's performance has degraded. By doing this, the model can be retrained at the right time, and the cost of the labeling service can be minimized.

   Source: Adapted from community discussion — TNT87

   References

   • Google Cloud documentation
   • Google Cloud documentation
   • Google developer documentation

   Next question

4. Question 44 · 364 in bank

   You work for a company that manages a ticketing platform for a large chain of cinemas. Customers use a mobile app to search for movies they’re interested in and purchase tickets in the app. Ticket purchase requests are sent to Pub/Sub and are processed with a Dataflow streaming pipeline configured to conduct the following steps: 1. Check for availability of the movie tickets at the selected cinema. 2. Assign the ticket price and accept payment. 3. Reserve the tickets at the selected cinema. 4. Send successful purchases to your database. Each step in this process has low latency requirements (less than 50 milliseconds). You have developed a logistic regression model with BigQuery ML that predicts whether offering a promo code for free popcorn increases the chance of a ticket purchase, and this prediction should be added to the ticket purchase process. You want to identify the simplest way to deploy this model to production while adding minimal latency. What should you do?

   • ARun batch inference with BigQuery ML every five minutes on each new set of tickets issued.
   • BExport your model in TensorFlow format, and add a tfx_bsl.public.beam.RunInference step to the Dataflow pipeline.
   • CExport your model in TensorFlow format, deploy it on Vertex AI, and query the prediction endpoint from your streaming pipeline.
   • DConvert your model with TensorFlow Lite (TFLite), and add it to the mobile app so that the promo code and the incoming request arrive together in Pub/Sub.

   Show answer

   Answer: D

   Explanation

   Question looks ambiguous! However considering some keywords like low latency and more importantly ML usage for maximising the ticket purchase requests using the promo code means that model embedded to the device looks more appropriate, however there are a lot of downsides to it like model management and upgrades but that does not seem to be the consideration here anyway. Just looking at low latency and ML to maximise the ticket sales, I will go with D as thats much simpler to implement

   Source: Adapted from community discussion — Krish6488

   Next question

5. Question 45 · 364 in bank

   You work on a team in a data center that is responsible for server maintenance. Your management team wants you to build a predictive maintenance solution that uses monitoring data to detect potential server failures. Incident data has not been labeled yet. What should you do first?

   • ATrain a time-series model to predict the machines’ performance values. Configure an alert if a machine’s actual performance values significantly differ from the predicted performance values.
   • BDevelop a simple heuristic (e.g., based on z-score) to label the machines’ historical performance data. Use this heuristic to monitor server performance in real time.
   • CDevelop a simple heuristic (e.g., based on z-score) to label the machines’ historical performance data. Train a model to predict anomalies based on this labeled dataset.
   • DHire a team of qualified analysts to review and label the machines’ historical performance data. Train a model based on this manually labeled dataset.

   Show answer

   Answer: B

   Explanation

   Changed to B, based on the comparison with #94, assuming that by “Use this heuristic to monitor server performance in real time” is meant to “first” test this heuristic for labelling in a Prod. environment, as a quick reality-check, before training a whole model on a roughly inaccurate labelled dataset.

   Source: Adapted from community discussion — M25

   References

   • Google developer documentation

   Next question

6. Question 46 · 364 in bank

   You built and manage a production system that is responsible for predicting sales numbers. Model accuracy is crucial, because the production model is required to keep up with market changes. Since being deployed to production, the model hasn't changed; however the accuracy of the model has steadily deteriorated. What issue is most likely causing the steady decline in model accuracy?

   • APoor data quality
   • BLack of model retraining
   • CToo few layers in the model for capturing information
   • DIncorrect data split ratio during model training, evaluation, validation, and test

   Show answer

   Answer: B

   Explanation

   I also think it is B - who is giving the "correct" answers to the questions? I feel like 4 out of 5 of them are incorrect.

   Source: Adapted from community discussion — sensev

   Next question

7. Question 47 · 364 in bank

   You work for a retailer that sells clothes to customers around the world. You have been tasked with ensuring that ML models are built in a secure manner. Specifically, you need to protect sensitive customer data that might be used in the models. You have identified four fields containing sensitive data that are being used by your data science team: AGE, IS_EXISTING_CUSTOMER, LATITUDE_LONGITUDE, and SHIRT_SIZE. What should you do with the data before it is made available to the data science team for training purposes?

   • ATokenize all of the fields using hashed dummy values to replace the real values.
   • BUse principal component analysis (PCA) to reduce the four sensitive fields to one PCA vector.
   • CCoarsen the data by putting AGE into quantiles and rounding LATITUDE_LONGTTUDE into single precision. The other two fields are already as coarse as possible.
   • DRemove all sensitive data fields, and ask the data science team to build their models using non-sensitive data.

   Show answer

   Answer: A

   Explanation

   B -> possible in general but not suitable in this case since you don't know AGE, IS_EXISTING_CUSTOMER, LATITUDE_LONGITUDE, and SHIRT_SIZE are the first components in PCA. C -> You are changing data which could be highly correlated with the output D -> like C explanation Answer 'A' uses hashing so you encript the data without losing relevant information

   Source: Adapted from community discussion — Scipione_

   References

   • Google Cloud documentation

   Next question

8. Question 48 · 364 in bank

   You work for a magazine publisher and have been tasked with predicting whether customers will cancel their annual subscription. In your exploratory data analysis, you find that 90% of individuals renew their subscription every year, and only 10% of individuals cancel their subscription. After training a NN Classifier, your model predicts those who cancel their subscription with 99% accuracy and predicts those who renew their subscription with 82% accuracy. How should you interpret these results?

   • AThis is not a good result because the model should have a higher accuracy for those who renew their subscription than for those who cancel their subscription.
   • BThis is not a good result because the model is performing worse than predicting that people will always renew their subscription.
   • CThis is a good result because predicting those who cancel their subscription is more difficult, since there is less data for this group.
   • DThis is a good result because the accuracy across both groups is greater than 80%.

   Show answer

   Answer: C

   Explanation

   Went with C: This is a good result because predicting those who cancel their subscription is more difficult, since there is less data for this group My Reason: "You have been tasked with predicting whether customers will cancel their annual subscription." And in that task you are getting 99% of accuracy

   Source: Adapted from community discussion — Voyager2

   Next question

9. Question 49 · 364 in bank

   You have built a model that is trained on data stored in Parquet files. You access the data through a Hive table hosted on Google Cloud. You preprocessed these data with PySpark and exported it as a CSV file into Cloud Storage. After preprocessing, you execute additional steps to train and evaluate your model. You want to parametrize this model training in Kubeflow Pipelines. What should you do?

   • ARemove the data transformation step from your pipeline.
   • BContainerize the PySpark transformation step, and add it to your pipeline.
   • CAdd a ContainerOp to your pipeline that spins a Dataproc cluster, runs a transformation, and then saves the transformed data in Cloud Storage.
   • DDeploy Apache Spark at a separate node pool in a Google Kubernetes Engine cluster. Add a ContainerOp to your pipeline that invokes a corresponding transformation job for this Spark instance.

   Show answer

   Answer: C

   Explanation

   Since the data is stored in Parquet format, it's more efficient to use Spark to transform it. Containerizing the PySpark transformation step and adding it to the pipeline may not be the optimal solution since it may require additional resources to run this container. Deploying Apache Spark at a separate node pool in a Google Kubernetes Engine cluster and adding a ContainerOp to invoke a corresponding transformation job for this Spark instance is also a possible solution, but it may require more setup and configuration. Using Dataproc can simplify this process since it's a fully managed service that simplifies running Apache Spark and Hadoop clusters. A ContainerOp can be added to the pipeline to spin up a Dataproc cluster, run the transformation using PySpark, and save the transformed data in Cloud Storage. This solution is more efficient since Dataproc can scale the cluster based on the size of the data and the complexity of the transformation.

   Source: Adapted from community discussion — tavva_prudhvi

   Next question

10. Question 50 · 364 in bank

    You have developed an ML model to detect the sentiment of users’ posts on your company's social media page to identify outages or bugs. You are using Dataflow to provide real-time predictions on data ingested from Pub/Sub. You plan to have multiple training iterations for your model and keep the latest two versions live after every run. You want to split the traffic between the versions in an 80:20 ratio, with the newest model getting the majority of the traffic. You want to keep the pipeline as simple as possible, with minimal management required. What should you do?

    • ADeploy the models to a Vertex AI endpoint using the traffic-split=0=80, PREVIOUS_MODEL_ID=20 configuration.
    • BWrap the models inside an App Engine application using the --splits PREVIOUS_VERSION=0.2, NEW_VERSION=0.8 configuration
    • CWrap the models inside a Cloud Run container using the REVISION1=20, REVISION2=80 revision configuration.
    • DImplement random splitting in Dataflow using beam.Partition() with a partition function calling a Vertex AI endpoint.

    Show answer

    Answer: A

    Explanation

    A. Deploy the models to a Vertex AI endpoint using the traffic-split=0=80, PREVIOUS_MODEL_ID=20 configuration. The recommended approach to achieve the desired outcome would be to deploy the ML models to a Vertex AI endpoint and configure the traffic splitting using the traffic-split parameter. The traffic-split parameter enables you to split traffic between multiple versions of a model based on a percentage split. In this case, the newest model should receive the majority of the traffic, which can be achieved by setting the traffic-split parameter to 0=80. The previous version of the model should receive the remaining 20% of the traffic, which can be achieved by setting the PREVIOUS_MODEL_ID parameter to 20.

    Source: Adapted from community discussion — TNT87

    References

    • Google Cloud documentation
    • Google Cloud documentation

    Next question

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

PMLE practice, question 41. 364 questions available.