Free Sample — 15 Practice Questions
Preview 15 of 333 questions from the Professional Machine Learning Engineer exam.
Try before you buy — purchase the full study guide for all 333 questions with answers and explanations.
Question 94
You built a deep learning-based image classification model by using on-premises data. You want to use Vertex AI to deploy the model to production. Due to security concerns, you cannot move your data to the cloud. You are aware that the input data distribution might change over time. You need to detect model performance changes in production. What should you do?
A. Use Vertex Explainable AI for model explainability. Configure feature-based explanations.
B. Use Vertex Explainable AI for model explainability. Configure example-based explanations.
C. Create a Vertex AI Model Monitoring job. Enable training-serving skew detection for your model.
D. Create a Vertex AI Model Monitoring job. Enable feature attribution skew and drift detection for your model.
Show Answer
Correct Answer: D
Explanation:
The goal is to detect changes in input data distribution and potential performance degradation in production without moving on‑premises training data to the cloud. Vertex Explainable AI (A, B) provides interpretability, not ongoing performance or drift monitoring. Training‑serving skew detection (C) requires access to training data as a baseline, which conflicts with the constraint of not moving data to the cloud. Feature attribution skew and drift detection in Vertex AI Model Monitoring (D) can detect data drift and behavioral changes in production using serving data and attributions, making it the most appropriate choice under the given security and monitoring requirements.
Question 40
You are building an application that extracts information from invoices and receipts. You want to implement this application with minimal custom code and training. What should you do?
A. Use the Cloud Vision API with TEXT_DETECTION type to extract text from the invoices and receipts, and use a pre-built natural language processing (NLP) model to parse the extracted text.
B. Use the Cloud Document AI API to extract information from the invoices and receipts.
C. Use Vertex AI Agent Builder with the pre-built Layout Parser model to extract information from the invoices and receipts.
D. Train an AutoML Natural Language model to classify and extract information from the invoices and receipts.
Show Answer
Correct Answer: B
Explanation:
The requirement is to extract structured information from invoices and receipts with minimal custom code and no training. Cloud Document AI provides pre-built, purpose-built processors (such as Invoice and Receipt Parsers) that automatically extract key fields like vendor, dates, line items, and totals. Other options either require additional parsing logic, custom model training, or are not specialized for structured document extraction.
Question 148
You work for a bank with strict data governance requirements. You recently implemented a custom model to detect fraudulent transactions. You want your training code to download internal data by using an API endpoint hosted in your project’s network. You need the data to be accessed in the most secure way, while mitigating the risk of data exfiltration. What should you do?
A. Enable VPC Service Controls for peerings, and add Vertex AI to a service perimeter.
B. Create a Cloud Run endpoint as a proxy to the data. Use Identity and Access Management (IAM) authentication to secure access to the endpoint from the training job.
C. Configure VPC Peering with Vertex AI, and specify the network of the training job.
D. Download the data to a Cloud Storage bucket before calling the training job.
Show Answer
Correct Answer: A
Explanation:
The requirement emphasizes maximum security and explicit mitigation of data exfiltration while allowing Vertex AI training to access internal data. VPC Service Controls are designed specifically to prevent data exfiltration by enforcing a service perimeter around Google-managed services. Adding Vertex AI to a service perimeter and enabling controls for peerings ensures that training jobs can only access approved internal resources and cannot move data outside the perimeter, even if credentials are compromised. The other options provide connectivity or access control, but they do not offer the same perimeter-based, exfiltration-focused protection.
Question 125
You have trained a model by using data that was preprocessed in a batch Dataflow pipeline. Your use case requires real-time inference. You want to ensure that the data preprocessing logic is applied consistently between training and serving. What should you do?
A. Perform data validation to ensure that the input data to the pipeline is the same format as the input data to the endpoint.
B. Refactor the transformation code in the batch data pipeline so that it can be used outside of the pipeline. Use the same code in the endpoint.
C. Refactor the transformation code in the batch data pipeline so that it can be used outside of the pipeline. Share this code with the end users of the endpoint.
D. Batch the real-time requests by using a time window and then use the Dataflow pipeline to preprocess the batched requests. Send the preprocessed requests to the endpoint.
Show Answer
Correct Answer: B
Explanation:
To ensure consistent preprocessing between training (batch Dataflow) and real-time serving, the same transformation logic must be reused. Refactoring the batch pipeline’s transformation code into reusable components and invoking that same code in the serving endpoint guarantees identical preprocessing. Data validation alone doesn’t enforce identical transformations, sharing code with end users is impractical and error-prone, and batching real-time requests through Dataflow introduces latency and undermines real-time inference.
Question 239
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?
A. Train 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.
B. Develop 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.
C. Develop 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.
D. Hire 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
Correct Answer: C
Explanation:
Incident data is unlabeled, so the first step is to create labels cheaply. A simple heuristic (e.g., z-score–based anomaly detection) can be used to label historical data, which then enables training a predictive model. This bootstraps a scalable predictive maintenance solution, whereas only monitoring with heuristics does not build a predictive model.
Question 228
You have been asked to build a model using a dataset that is stored in a medium-sized (~10 GB) BigQuery table. You need to quickly determine whether this data is suitable for model development. You want to create a one-time report that includes both informative visualizations of data distributions and more sophisticated statistical analyses to share with other ML engineers on your team. You require maximum flexibility to create your report. What should you do?
A. Use Vertex AI Workbench user-managed notebooks to generate the report.
B. Use the Google Data Studio to create the report.
C. Use the output from TensorFlow Data Validation on Dataflow to generate the report.
D. Use Dataprep to create the report.
Show Answer
Correct Answer: A
Explanation:
You need a one-time exploratory report with both rich visualizations and sophisticated statistical analysis, and you want maximum flexibility. Vertex AI Workbench user-managed notebooks let you read directly from BigQuery and use Python libraries (pandas, matplotlib/seaborn, scipy, statsmodels) to freely combine custom analysis, plots, and narrative. Looker Studio is geared toward dashboards with limited statistical depth, TFDV is specialized for schema/validation rather than flexible EDA, and Dataprep focuses on data wrangling rather than advanced analysis and reporting.
Question 128
You are building a predictive maintenance model to preemptively detect part defects in bridges. You plan to use high definition images of the bridges as model inputs. You need to explain the output of the model to the relevant stakeholders so they can take appropriate action. How should you build the model?
A. Use scikit-learn to build a tree-based model, and use SHAP values to explain the model output.
B. Use scikit-learn to build a tree-based model, and use partial dependence plots (PDP) to explain the model output.
C. Use TensorFlow to create a deep learning-based model, and use Integrated Gradients to explain the model output.
D. Use TensorFlow to create a deep learning-based model, and use the sampled Shapley method to explain the model output.
Show Answer
Correct Answer: C
Explanation:
High‑definition bridge images are best handled by deep learning (e.g., CNNs built with TensorFlow), which can automatically learn spatial and visual features relevant to defects. Integrated Gradients is specifically designed to explain deep neural network predictions by attributing importance to input pixels, producing intuitive visual explanations (e.g., heatmaps) that stakeholders can understand and act upon.
Question 254
You work for a large social network service provider whose users post articles and discuss news. Millions of comments are posted online each day, and more than 200 human moderators constantly review comments and flag those that are inappropriate. Your team is building an ML model to help human moderators check content on the platform. The model scores each comment and flags suspicious comments to be reviewed by a human. Which metric(s) should you use to monitor the model’s performance?
A. Number of messages flagged by the model per minute
B. Number of messages flagged by the model per minute confirmed as being inappropriate by humans.
C. Precision and recall estimates based on a random sample of 0.1% of raw messages each minute sent to a human for review
D. Precision and recall estimates based on a sample of messages flagged by the model as potentially inappropriate each minute
Show Answer
Correct Answer: C
Explanation:
To monitor model performance you need metrics that reflect correctness, not just volume. Precision and recall are appropriate because they capture false positives and false negatives. Estimating them requires an unbiased sample of all messages, not only those flagged by the model. Sampling only flagged messages (option D) allows estimating precision but cannot estimate recall because false negatives are never reviewed. A random sample of raw messages reviewed by humans enables identifying both missed inappropriate content and correctly flagged content, making option C the correct choice.
Question 90
You are developing a training pipeline for a new XGBoost classification model based on tabular data. The data is stored in a BigQuery table. You need to complete the following steps:
1. Randomly split the data into training and evaluation datasets in a 65/35 ratio
2. Conduct feature engineering
3. Obtain metrics for the evaluation dataset
4. Compare models trained in different pipeline executions
How should you execute these steps?
A. 1. Using Vertex AI Pipelines, add a component to divide the data into training and evaluation sets, and add another component for feature engineering. 2. Enable autologging of metrics in the training component. 3. Compare pipeline runs in Vertex AI Experiments.
B. 1. Using Vertex AI Pipelines, add a component to divide the data into training and evaluation sets, and add another component for feature engineering. 2. Enable autologging of metrics in the training component. 3. Compare models using the artifacts’ lineage in Vertex ML Metadata.
C. 1. In BigQuery ML, use the CREATE MODEL statement with BOOSTED_TREE_CLASSIFIER as the model type and use BigQuery to handle the data splits. 2. Use a SQL view to apply feature engineering and train the model using the data in that view. 3. Compare the evaluation metrics of the models by using a SQL query with the ML.TRAINING_INFO statement.
D. 1. In BigQuery ML, use the CREATE MODEL statement with BOOSTED_TREE_CLASSIFIER as the model type and use BigQuery to handle the data splits. 2. Use ML TRANSFORM to specify the feature engineering transformations and tram the model using the data in the table. 3. Compare the evaluation metrics of the models by using a SQL query with the ML.TRAINING_INFO statement.
Show Answer
Correct Answer: A
Explanation:
The requirements call for an end-to-end training pipeline with data splitting, feature engineering, metric tracking, and comparison across different pipeline executions. Vertex AI Pipelines support modular components for splitting data and feature engineering. Enabling autologging in the training step captures evaluation metrics automatically. Vertex AI Experiments is specifically designed to compare metrics and parameters across multiple pipeline runs, making it the correct tool for model comparison. BigQuery ML options are less suitable for orchestrated pipelines and cross-run experiment comparison, and ML Metadata lineage is not intended for direct model performance comparison.
Question 27
You are training a large-scale deep learning model on a Cloud TPU. While monitoring the training progress through Tensorboard, you observe that the TPU utilization is consistently low and there are delays between the completion of one training step and the start of the next step. You want to improve TPU utilization and overall training performance. How should you address this issue?
A. Apply tf.data.Detaset.map with vectorized operations and parallelization.
B. Use tf.data.Detaset.interleave with multiple data sources.
C. Use tf.data.Detaset.cache on the dataset after the first epoch.
D. Implement tf.data.Detaset.prefetch in the data pipeline.
Show Answer
Correct Answer: D
Explanation:
Low TPU utilization with gaps between steps indicates the input pipeline is a bottleneck. tf.data.Dataset.prefetch overlaps data input/preprocessing with TPU execution so the next batch is ready when the current step finishes, reducing idle time and improving utilization. The other options can help in some cases, but prefetch directly addresses step-to-step delays.
Question 136
You work as an analyst at a large banking firm. You are developing a robust scalable ML pipeline to tram several regression and classification models. Your primary focus for the pipeline is model interpretability. You want to productionize the pipeline as quickly as possible. What should you do?
A. Use Tabular Workflow for Wide & Deep through Vertex AI Pipelines to jointly train wide linear models and deep neural networks
B. Use Google Kubernetes Engine to build a custom training pipeline for XGBoost-based models
C. Use Tabular Workflow for TabNet through Vertex AI Pipelines to train attention-based models
D. Use Cloud Composer to build the training pipelines for custom deep learning-based models
Show Answer
Correct Answer: C
Explanation:
The requirements emphasize strong model interpretability and rapid productionization for tabular regression and classification. Vertex AI Tabular Workflows with TabNet are purpose-built for this: they are managed, scalable, and quick to deploy, and TabNet provides inherent interpretability via its sequential attention mechanism with feature attributions. The other options either emphasize less interpretable models, require more custom engineering, or are not optimized for fast, managed ML pipelines.
Question 150
You work for a bank. You have created a custom model to predict whether a loan application should be flagged for human review. The input features are stored in a BigQuery table. The model is performing well, and you plan to deploy it to production. Due to compliance requirements the model must provide explanations for each prediction. You want to add this functionality to your model code with minimal effort and provide explanations that are as accurate as possible. What should you do?
A. Create an AutoML tabular model by using the BigQuery data with integrated Vertex Explainable AI.
B. Create a BigQuery ML deep neural network model and use the ML.EXPLAIN_PREDICT method with the num_integral_steps parameter.
C. Upload the custom model to Vertex AI Model Registry and configure feature-based attribution by using sampled Shapley with input baselines.
D. Update the custom serving container to include sampled Shapley-based explanations in the prediction outputs.
Show Answer
Correct Answer: C
Explanation:
You already have a well-performing custom model and need per-prediction explanations with minimal additional effort and maximum fidelity. Vertex AI Explainable AI supports feature-based attribution (including sampled Shapley with baselines) for custom models without modifying model code. Uploading the model to Vertex AI Model Registry and enabling explanations provides accurate, compliant explanations at serving time. Rebuilding the model (A, B) adds unnecessary effort, and implementing Shapley inside the serving container (D) requires significantly more custom work than using Vertex AI’s built-in XAI.
Question 212
You recently developed a deep learning model using Keras, and now you are experimenting with different training strategies. First, you trained the model using a single GPU, but the training process was too slow. Next, you distributed the training across 4 GPUs using tf.distribute.MirroredStrategy (with no other changes), but you did not observe a decrease in training time. What should you do?
A. Distribute the dataset with tf.distribute.Strategy.experimental_distribute_dataset
B. Create a custom training loop.
C. Use a TPU with tf.distribute.TPUStrategy.
D. Increase the batch size.
Show Answer
Correct Answer: D
Explanation:
With tf.distribute.MirroredStrategy, the model is replicated on each GPU and the global batch is split across replicas. If you keep the same batch size as single‑GPU training, each GPU receives a smaller per‑replica batch, leading to poor utilization and little or no speedup. Increasing the global batch size (typically proportional to the number of GPUs) allows each GPU to do more work in parallel and amortizes communication overhead, which is the standard fix when multi‑GPU training shows no performance gain.
Question 295
As the lead ML Engineer for your company, you are responsible for building ML models to digitize scanned customer forms. You have developed a TensorFlow model that converts the scanned images into text and stores them in Cloud Storage. You need to use your ML model on the aggregated data collected at the end of each day with minimal manual intervention. What should you do?
A. Use the batch prediction functionality of AI Platform.
B. Create a serving pipeline in Compute Engine for prediction.
C. Use Cloud Functions for prediction each time a new data point is ingested.
D. Deploy the model on AI Platform and create a version of it for online inference.
Show Answer
Correct Answer: A
Explanation:
The data is processed in aggregated batches at the end of each day, not in real time. AI Platform batch prediction is designed for exactly this scenario: running large-scale, offline inference on stored data with minimal infrastructure management and manual intervention. Online prediction (D, C) is unnecessary, and building custom serving on Compute Engine (B) increases operational overhead.
Question 4
You built a custom Vertex AI pipeline job that preprocesses images and trains an object detection model. The pipeline currently uses 1 n1-standard-8 machine with 1 NVIDIA Tesla V100 GPU. You want to reduce the model training time without compromising model accuracy. What should you do?
A. Reduce the number of layers in your object detection model.
B. Train the same model on a stratified subset of your dataset.
C. Update the WorkerPoolSpec to use a machine with 24 vCPUs and 1 NVIDIA Tesla V100 GPU.
D. Update the WorkerPoolSpec to use a machine with 24 vCPUs and 3 NVIDIA Tesla V100 GPUs.
Show Answer
Correct Answer: D
Explanation:
To reduce training time without sacrificing accuracy, you need to increase compute resources rather than change the model or data. Options A and B would likely reduce model capacity or effective data, harming accuracy. Simply increasing CPU cores (option C) offers limited benefit because object detection training is GPU-bound. Using a machine with multiple GPUs (option D) enables parallel training and significantly speeds up training while preserving the same model and dataset.