Google ML Engineer Advanc - Free Practice Quiz

Question 1 of 60

A Google Professional ML Engineer 2023 exam candidate must design a training pipeline that guarantees consistent feature transformations between offline training and online serving for a fraud detection model. Which architecture MOST effectively prevents training-serving skew?

A Computing features in separate Python scripts maintained independently by the data science and engineering teams B Applying normalization only to training data and hardcoding serving-time scaling parameters C Defining all feature transformations as TFX Transform components that generate both the transformed training data and the serving graph, then deploying the serving graph alongside the model D Caching the last batch of training features and reusing them at serving time

1 / 60

Question 2 of 60

A Google Professional ML Engineer 2023 exam candidate is designing a pipeline for real-time anomaly detection on IoT sensor streams. Which end-to-end architecture on Google Cloud BEST meets sub-second latency and scalability requirements?

A Cloud Storage batch upload â†’ Vertex AI Batch Prediction â†’ Cloud Scheduler reporting B BigQuery scheduled queries â†’ Cloud Functions â†’ email alerts C Cloud Pub/Sub â†’ Cloud Dataflow (streaming) â†’ Vertex AI Online Prediction endpoint â†’ anomaly alerts via Pub/Sub or alerting D Cloud Bigtable â†’ Cloud Dataproc Spark â†’ BigQuery export

2 / 60

Question 3 of 60

The Google Professional ML Engineer 2023 exam covers designing ML systems for high-stakes decisions. Which evaluation practice MOST ensures a credit risk model does not systematically disadvantage a protected class?

A Reporting overall AUC-ROC only and submitting for production B Using the highest-performing model regardless of demographic performance differences C Measuring false positive and false negative rates separately for each protected demographic group and applying fairness constraints during training D Relying on the model's cross-entropy loss as a proxy for fairness

3 / 60

Question 4 of 60

A Google Professional ML Engineer 2023 exam candidate is tasked with reducing the cost of a Vertex AI training job that requires 48 hours on 8 Ã— A100 GPUs. Which optimization MOST reduces cost without sacrificing model quality?

A Switching to a CPU-only training cluster B Removing batch normalization layers to speed up each step C Mixed-precision training (float16/bfloat16) reducing memory usage and increasing throughput, combined with gradient checkpointing to allow larger effective batch sizes D Disabling early stopping to ensure full training

4 / 60

Question 5 of 60

On the Google Professional ML Engineer 2023 exam, a data scientist observes that their gradient boosted model achieves 99% training accuracy but only 62% test accuracy. Which combination of interventions is MOST likely to address this?

A Increasing tree depth and removing all regularization to capture more patterns B Switching to a linear model and increasing the number of features C Collecting more training data, reducing tree depth, lowering the learning rate, and enabling subsampling regularization D Applying one-hot encoding to all features and re-running with the same hyperparameters

5 / 60

Question 6 of 60

On the Google Professional ML Engineer 2023 exam, which strategy MOST effectively addresses label noise in a large training dataset without discarding data?

A Replacing all labels with model pseudo-labels before final training B Removing any example where the model's prediction differs from the label C Training with label noise-robust loss functions (e.g., symmetric cross-entropy) or learning with noisy labels (LNL) techniques that down-weight potentially mislabeled examples D Increasing model capacity to absorb label noise

6 / 60

Question 7 of 60

On the Google Professional ML Engineer 2023 exam, an engineer needs to serve a 13B-parameter language model with p99 latency under 200ms on Vertex AI. Which approach MOST likely meets this requirement?

A Deploying the full float32 model on CPU instances B Serving only from the model's first 3 layers to reduce inference depth C Using 8-bit quantization, optimizing with TensorRT or vLLM, and deploying on multi-GPU A100 instances with autoscaling D Batching all requests into hourly offline jobs

7 / 60

Question 8 of 60

The Google Professional ML Engineer 2023 exam covers 'Shapley values' for model explainability. Which limitation of Shapley-based explanations is MOST important to communicate to stakeholders?

A Shapley values only work for linear models B Shapley values cannot be computed for classification models C Shapley values are computationally expensive for large feature sets and explain predictions at the local (per-instance) level â€” they do not represent global model behavior for all inputs D Shapley values explain all model errors perfectly

8 / 60

Question 9 of 60

According to the Google Professional ML Engineer 2023 exam, a production ML model's AUC-ROC has degraded from 0.92 to 0.74 over 6 months. Vertex AI Model Monitoring alerts show significant feature drift in 3 key inputs. Which response plan is MOST appropriate?

A Immediately roll back to a previous model version and disable monitoring B Ignore the drift since AUC-ROC above 0.70 is generally acceptable C Investigate the drifting features for root cause, refresh the training dataset with recent labeled examples, retrain the model, and validate before redeployment via the existing Vertex AI Pipeline D Increase the model's decision threshold to compensate for degraded discrimination

9 / 60

Question 10 of 60

According to the Google Professional ML Engineer 2023 exam, which approach MOST effectively validates that a newly trained model is a genuine improvement before production deployment?

A Comparing only training set metrics between the old and new models B Deploying the new model as soon as training completes without comparison C Statistical significance testing of the new model's offline evaluation metrics versus the current production model using a held-out evaluation dataset, gated by minimum acceptable improvement thresholds D Sending the new model directly to 100% of production traffic

10 / 60

Question 11 of 60

According to the Google Professional ML Engineer 2023 exam, which technique MOST effectively handles long-range dependencies in a document classification model trained on variable-length text inputs?

A Using fixed-length input truncation with a simple bag-of-words model B Applying character-level CNN without positional encoding C Fine-tuning a pretrained Transformer (e.g., BERT or Longformer for documents > 512 tokens) with positional encoding capturing sequence structure D Using a BiLSTM without attention

11 / 60

Question 12 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST effectively makes a recommendation model robust to popularity bias (where popular items dominate predictions)?

A Removing all items with fewer than 1,000 interactions from training B Sorting recommendations by item ID alphabetically C Applying inverse popularity weighting to training samples so rare interactions contribute more, and evaluating recommendations on diversity and coverage metrics in addition to accuracy D Using accuracy as the sole evaluation metric

12 / 60

Question 13 of 60

The Google Professional ML Engineer 2023 exam addresses serving large language models (LLMs) on Vertex AI. Which serving configuration MOST reduces per-request cost while maintaining acceptable latency for a high-volume summarization API?

A Using a single CPU-based instance to minimize GPU costs B Deploying the full 70B parameter model on TPU v4 pods with no quantization C Applying 4-bit quantization, enabling request batching, and deploying on GPU instances with autoscaling â€” balanced for throughput and cost D Increasing batch size to maximum and serving synchronously

13 / 60

Question 14 of 60

The Google Professional ML Engineer 2023 exam covers serving ML models at the edge. Which combination of tools is MOST appropriate for converting a Vertex AI trained TensorFlow model to run on mobile devices?

A Using PyTorch Mobile since TensorFlow models cannot be converted for mobile B Deploying the full SavedModel directly to a mobile app without conversion C TensorFlow Lite (TFLite) conversion with quantization, validated on device benchmarks, then distributed via an OTA model update mechanism D Converting the model to ONNX and serving via a cloud endpoint called by the mobile app

14 / 60

Question 15 of 60

The Google Professional ML Engineer 2023 exam asks about multi-modal ML. Which Vertex AI capability MOST supports building a model that jointly processes both image and text inputs?

A Using separate models for images and text and averaging their outputs manually B Using BigQuery ML which natively supports multi-modal inputs C Training a joint multi-modal architecture (e.g., a Vision-Language Model) using Vertex AI Custom Training with a custom multi-modal data pipeline D Restricting the model to single-modality inputs for simplicity

15 / 60

Question 16 of 60

According to the Google Professional ML Engineer 2023 exam, which Vertex AI capability MOST efficiently enables parameter-efficient fine-tuning of a foundation model on a small domain-specific dataset?

A Low-Rank Adaptation (LoRA) or adapter tuning â€” fine-tuning a small number of additional parameters while keeping most pretrained weights frozen, dramatically reducing compute requirements B Full fine-tuning of all model parameters on the small dataset C Using a bag-of-words model instead of the foundation model D Retraining the foundation model from random weights on the domain dataset

16 / 60

Question 17 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST effectively evaluates fairness of a binary classification model used in a loan approval system?

A Computing equalized odds and demographic parity difference disaggregated by protected attributes, and using Vertex AI Explainability to audit feature attributions for bias signals B Relying on the model vendor's fairness certification C Using accuracy as the sole metric since it reflects overall correctness D Reporting only the overall AUC-ROC on the full test set

17 / 60

Question 18 of 60

On the Google Professional ML Engineer 2023 exam, which approach to handling severe class imbalance (1:1000 ratio) in a classification task is MOST likely to produce a well-calibrated model?

A Combining oversampling of the minority class, undersampling of the majority, class-weighted loss, and calibrating the final model's probability outputs using Platt scaling or isotonic regression B Using accuracy as the optimization metric since it captures all classes C Removing all majority class examples until a 1:1 ratio is achieved D Using a threshold of 0.5 for classification without any imbalance handling

18 / 60

Question 19 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST effectively enables a Vertex AI training pipeline to recover from hardware failures mid-run without losing all progress?

A Implementing checkpointing â€” saving model weights and optimizer state to Cloud Storage at regular intervals and configuring the training job to resume from the latest checkpoint B Storing the full training dataset in Cloud Bigtable for fast re-access C Relying on the CSP to automatically restart jobs without any configuration D Restarting from random weight initialization after each failure

19 / 60

Question 20 of 60

The Google Professional ML Engineer 2023 exam describes 'concept drift' versus 'data drift'. Which scenario BEST illustrates concept drift specifically?

A Consumer spending patterns fundamentally change during a recession, so the relationship between historical transaction features and the fraud label no longer holds B The volume of incoming data triples during peak season C Input feature schema changes due to upstream pipeline modification D The distribution of raw input features shifts while the relationship between inputs and labels remains stable

20 / 60

Question 21 of 60

A Google Professional ML Engineer 2023 exam candidate is designing a multi-region ML serving architecture that must minimize prediction latency globally. Which Vertex AI deployment pattern BEST achieves this?

A Deploying Vertex AI Online Prediction endpoints in multiple Google Cloud regions and using global load balancing with latency-based routing to direct requests to the nearest endpoint B Serving all predictions from a batch export refreshed hourly C Routing all users to the nearest Cloud CDN cache for prediction results D Deploying a single endpoint in us-central1 and relying on Google's global network to minimize latency

21 / 60

Question 22 of 60

The Google Professional ML Engineer 2023 exam asks about designing an ML system that must comply with GDPR's right to erasure. Which architecture MOST effectively enables erasure without full model retraining?

A Cryptographic erasure of the individual's data encryption keys in the feature store, combined with machine unlearning techniques applied to model weights, and retraining on erased datasets during the next scheduled pipeline run B Deleting the user's raw data from Cloud Storage only C Responding to erasure requests by deploying a new randomly initialized model D Anonymizing the user's data by replacing all PII with random values

22 / 60

Question 23 of 60

According to the Google Professional ML Engineer 2023 exam, which combination of Vertex AI components MOST effectively supports a regulated ML model's 'right to explanation' requirement?

A Vertex AI Explainability (SHAP/Integrated Gradients) with per-prediction feature attribution logs, plus Vertex AI Experiments for training lineage B Vertex AI Workbench and Cloud Logging combined C Cloud Monitoring latency dashboards and BigQuery audit logs D Vertex AI Feature Store and Cloud Dataflow combined

23 / 60

Question 24 of 60

On the Google Professional ML Engineer 2023 exam, a team trains a model on data from a cloud-based simulation environment to be deployed in a physical robotic system. The model performs well in simulation but fails in the real world. Which phenomenon MOST explains this?

A The simulation-to-reality (Sim2Real) gap â€” differences between simulated and real-world data distributions cause poor transfer B The model used too many hidden layers C The simulation used float32 precision instead of float16 D The simulation had too much labeled data

24 / 60

Question 25 of 60

The Google Professional ML Engineer 2023 exam covers federated learning. In which scenario is federated learning MOST appropriate?

A When training data is distributed across user devices or organizations and cannot be centralized due to privacy or regulatory constraints B When training data is small enough to fit in a single BigQuery table C When the training dataset is publicly available with no privacy requirements D When all training data is centralized in a single Google Cloud project

25 / 60

Question 26 of 60

According to the Google Professional ML Engineer 2023 exam, which approach MOST effectively reduces Vertex AI training cost for a large-scale NLP model without sacrificing final model quality?

A Pretraining on a small domain-specific dataset first, then fine-tuning a smaller model using knowledge distillation from a larger pretrained foundation model â€” reducing GPU hours while achieving near-teacher accuracy B Training the model from random weights on the full dataset with maximum GPU resources C Using exclusively preemptible TPU pods without checkpointing D Reducing training dataset size by 50% before starting training

26 / 60

Question 27 of 60

The Google Professional ML Engineer 2023 exam covers 'prompt engineering' for LLM-based applications. Which technique MOST reliably improves the accuracy of an LLM's structured output (e.g., JSON extraction) without fine-tuning?

A Few-shot prompting with well-structured examples of the desired input-output format, combined with output format constraints B Using a single-word system prompt C Feeding the model unformatted free-text instructions D Reducing the model's temperature to zero and relying on greedy decoding alone

27 / 60

Question 28 of 60

According to the Google Professional ML Engineer 2023 exam, which technique MOST effectively enables an ML model to refuse to make predictions when input data falls outside the training distribution?

A Out-of-distribution (OOD) detection using methods like Mahalanobis distance, energy-based OOD scoring, or conformal prediction â€” flagging inputs where the model's confidence is unreliable B Using the model's softmax probabilities alone as an OOD indicator C Increasing model complexity until OOD inputs are classified correctly D Setting a fixed confidence threshold of 0.5 for all predictions

28 / 60

Question 29 of 60

On the Google Professional ML Engineer 2023 exam, a continuous training pipeline has been deployed with Vertex AI Pipelines. Which additional component MOST completes the MLOps feedback loop?

A Vertex AI Model Monitoring with automated retraining triggers â€” completing the loop from production drift detection through data refresh, retraining, evaluation, and conditional model promotion B Deploying the model to a second region for redundancy C Exporting training logs to Cloud Storage for archival D A manual weekly review meeting to assess model quality

29 / 60

Question 30 of 60

The Google Professional ML Engineer 2023 exam covers 'multi-objective optimization' in ML. Which scenario MOST requires this approach?

A Optimizing a model to simultaneously maximize prediction accuracy while minimizing inference latency and memory footprint â€” requiring Pareto-optimal trade-off analysis B Training a model on a single GPU with maximum batch size C Selecting the optimal regularization lambda using grid search D Choosing between ReLU and sigmoid activation functions

30 / 60

Question 31 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST effectively evaluates the business impact of deploying a new ML model version in production?

A Comparing the model's offline AUC-ROC against the previous version B Running an online A/B test with randomized user assignment, measuring business KPIs (conversion rate, revenue per user, churn) for both versions, and using statistical significance testing C Using only the model training loss as the business metric D Reviewing user complaints submitted within 7 days of deployment

31 / 60

Question 32 of 60

The Google Professional ML Engineer 2023 exam covers responsible AI in generative models. Which risk is MOST specific to deploying a large language model (LLM) in a customer-facing application?

A The LLM requires more storage than traditional ML models B Hallucination â€” the LLM generating confident but factually incorrect or fabricated information, misleading users C LLMs are unable to process text inputs longer than 128 tokens D The LLM produces deterministic outputs that are easy to audit

32 / 60

Question 33 of 60

According to the Google Professional ML Engineer 2023 exam, which technique MOST effectively reduces catastrophic forgetting when incrementally fine-tuning a pretrained language model on a new domain?

A Training on only the new domain data with a high learning rate and no regularization B Elastic Weight Consolidation (EWC) or replay methods that retain a subset of old training data to penalize changes to weights critical for previous tasks C Completely retraining the model from random weights on the combined old and new data D Freezing all pretrained weights and training only a single linear output layer

33 / 60

Question 34 of 60

On the Google Professional ML Engineer 2023 exam, which pattern MOST effectively enables zero-downtime model updates for a production Vertex AI Online Prediction endpoint serving millions of requests per day?

A Deploying the new model version by replacing the existing deployment, accepting a brief outage B Traffic splitting: deploy the new model version alongside the current version, route 1% of traffic to the new version, gradually increase traffic while monitoring latency and accuracy, then complete cutover C Deploying the new model to a different endpoint URL and updating client configuration separately D Requiring all clients to handle connection errors during deployment

34 / 60

Question 35 of 60

According to the Google Professional ML Engineer 2023 exam, which Vertex AI serving feature MOST reduces cold-start latency for a model endpoint that receives intermittent, bursty traffic?

A Enabling accelerator types on the endpoint B Configuring a minimum number of replicas (min_replica_count > 0) to keep warm instances always available C Disabling autoscaling to maintain fixed replicas D Using a larger machine type for each prediction replica

35 / 60

Question 36 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST effectively enables a Vertex AI model to serve predictions for user queries that include recently created entities (e.g., new products) not present in training data?

A Retraining the model daily on the full historical + new data B Retrieval-Augmented Generation (RAG) â€” combining the generative model with real-time retrieval from an up-to-date knowledge store (e.g., Vertex AI Vector Search) to ground responses in current facts C Using the model's softmax output for the 'unknown' class D Caching the model's predictions for known entities and ignoring new ones

36 / 60

Question 37 of 60

The Google Professional ML Engineer 2023 exam asks about designing a real-time recommendation system for a streaming platform. Which architecture MOST efficiently serves personalized recommendations at scale?

A Generating recommendations in a weekly batch job and displaying the same list all week B A two-stage retrieval-and-ranking pipeline: approximate nearest-neighbor retrieval (e.g., Vertex AI Vector Search) to generate candidates, then a lightweight ranking model served via Vertex AI Online Prediction C Querying the Vertex AI endpoint synchronously for each page load with the full candidate item set D Using a rule-based system that ranks items by global popularity only

37 / 60

Question 38 of 60

According to the Google Professional ML Engineer 2023 exam, which approach MOST effectively enables a team to detect silent data pipeline failures that corrupt ML feature values before they affect model predictions?

A Reviewing feature values in notebooks manually each week B Integrating TensorFlow Data Validation (TFDV) schema checks into the Vertex AI Pipeline to automatically detect anomalies (missing values, distribution shifts, schema violations) in each data batch before it reaches model training or serving C Relying on model accuracy degradation alerts as the primary data quality signal D Logging all raw input records to Cloud Storage for periodic human review

38 / 60

Question 39 of 60

The Google Professional ML Engineer 2023 exam covers 'data-centric AI'. Which practice MOST embodies this paradigm?

A Focusing primarily on architectural innovations and hyperparameter tuning B Investing in dataset curation â€” systematic error analysis, consistent labeling guidelines, targeted data collection for underperforming slices, and iterative dataset improvement C Maximizing model parameter count as the primary quality driver D Relying on automated data generation pipelines without human oversight

39 / 60

Question 40 of 60

According to the Google Professional ML Engineer 2023 exam, which strategy MOST effectively prevents a Vertex AI AutoML model from selecting features that would create discriminatory model behavior?

A Allowing AutoML to automatically select all available features without review B Explicitly excluding protected attributes and their proxies (e.g., zip code as a race proxy) from the feature set before AutoML training, then auditing the selected features and evaluating disaggregated performance metrics C Disabling all automated feature selection D Using only features with the highest correlation to the target variable

40 / 60

Question 41 of 60

On the Google Professional ML Engineer 2023 exam, which approach to ML pipeline testing MOST rigorously validates that a new training pipeline version will produce a deployable model?

A Running the pipeline on a subset of data and accepting any model that completes training B Unit testing each pipeline component, integration testing the full pipeline on a representative data sample, and gate-checking the model against minimum performance thresholds before promotion C Testing only the data ingestion step and trusting the rest D Skipping automated tests in favor of human review before deployment

41 / 60

Question 42 of 60

The Google Professional ML Engineer 2023 exam identifies which technique as MOST effective for improving generalization in a deep learning model trained on a limited labeled dataset?

A Training the model from scratch with maximum depth and width B Fine-tuning a pretrained foundation model using transfer learning, combined with data augmentation and dropout regularization â€” maximizing information extracted from limited labeled data C Removing all dropout and regularization to allow maximum fitting of the limited data D Using only 50% of available labeled data to force the model to generalize

42 / 60

Question 43 of 60

On the Google Professional ML Engineer 2023 exam, which ML system design pattern MOST effectively addresses the challenge of serving different user populations with distinct data distributions using a single model?

A Training a separate model for each user population from scratch B Conditioning model behavior on user segment embeddings or metadata features so the model learns population-specific patterns from a shared architecture C Using the global population model for all users equally D Filtering training data to the largest population only

43 / 60

Question 44 of 60

The Google Professional ML Engineer 2023 exam identifies which advanced Vertex AI Vizier configuration as MOST effective for optimizing a non-convex objective with multiple local minima?

A Grid search over a coarse parameter grid B Evolutionary or population-based search strategies (e.g., CMA-ES) configured in Vertex AI Vizier for exploring non-convex hyperparameter landscapes C Gradient-based optimization of hyperparameters D Random search with uniform sampling over all parameters

44 / 60

Question 45 of 60

According to the Google Professional ML Engineer 2023 exam, which regularization technique is MOST effective for sparse feature spaces with many irrelevant features?

A L2 regularization (Ridge) B L1 regularization (Lasso) which drives irrelevant feature weights to exactly zero, performing implicit feature selection C Batch normalization D Dropout applied to the input layer

45 / 60

Question 46 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST rigorously ensures that a deployed Vertex AI model does not produce harmful or biased outputs for a public-facing application?

A Restricting the model to returning only the top-1 prediction without confidence scores B Publishing the model without testing since bias only emerges from user feedback C Relying exclusively on the model's training loss as a proxy for fairness D Pre-deployment: bias evaluation with disaggregated metrics, red team adversarial testing, output filtering; post-deployment: continuous monitoring with human-in-the-loop review of flagged outputs and model update cycle

46 / 60

Question 47 of 60

According to the Google Professional ML Engineer 2023 exam, which approach MOST effectively benchmarks a new Vertex AI model endpoint for production readiness under realistic traffic patterns?

A Testing the endpoint with a single sequential request and measuring latency B Reviewing the training job logs for GPU utilization C Comparing training accuracy across three model versions D Load testing the endpoint with realistic concurrency levels using tools like Locust or Apache Benchmark, measuring p50/p95/p99 latency and error rates against SLO thresholds

47 / 60

Question 48 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST effectively monitors a Vertex AI model serving financial predictions for signs of adversarial manipulation by external actors?

A Reviewing monthly summary reports of model accuracy B Logging only prediction latency and HTTP status codes C Relying on user reports of incorrect predictions D Implementing anomaly detection on input feature distributions combined with adversarial example detectors, and alerting when inputs statistically resemble known adversarial patterns

48 / 60

Question 49 of 60

The Google Professional ML Engineer 2023 exam covers 'model cards' as a tool for:

A Storing trained model weights in a compressed format B Defining the serving infrastructure requirements for a model C Tracking hyperparameter configurations across training experiments D Providing a structured document that communicates intended uses, performance characteristics, and ethical considerations of a deployed ML model

49 / 60

Question 50 of 60

The Google Professional ML Engineer 2023 exam covers distributed training strategies. Which strategy is MOST appropriate for training a model too large to fit in a single accelerator's memory?

A Reducing the model to fit on one device by removing layers B Data parallelism â€” replicating the full model on each device and splitting data batches C Gradient accumulation with a single device D Model parallelism â€” splitting model layers across multiple accelerators so each device holds only a portion of the model

50 / 60

Question 51 of 60

The Google Professional ML Engineer 2023 exam identifies which scenario as MOST appropriate for Vertex AI Vector Search (formerly Matching Engine)?

A Running batch SQL analytics on structured tabular data B Running hyperparameter search across thousands of model configurations C Storing raw training datasets for periodic ML model retraining D Nearest-neighbor retrieval of relevant documents or items from a high-dimensional embedding space for semantic search or recommendation

51 / 60

Question 52 of 60

According to the Google Professional ML Engineer 2023 exam, which design choice MOST ensures long-term maintainability of a production ML system?

A Using proprietary frameworks that maximize performance on current hardware B Minimizing documentation to keep the codebase small C Hard-coding all model hyperparameters into the serving container D Modular pipeline design with versioned components, reproducible training via IaC, automated testing gates, and documented model cards â€” enabling future engineers to understand, test, and improve the system

52 / 60

Question 53 of 60

On the Google Professional ML Engineer 2023 exam, a Vertex AI Batch Prediction job is running significantly over its time budget. Which optimization MOST likely reduces job duration?

A Increasing the number of output shards without changing worker configuration B Removing all preprocessing steps from the pipeline C Switching from JSON to CSV output format D Increasing the number of compute replicas, enabling accelerators (GPUs/TPUs) for compute-bound models, and pre-sharding input data to maximize parallelism

53 / 60

Question 54 of 60

According to the Google Professional ML Engineer 2023 exam, which monitoring strategy MOST effectively detects when a Vertex AI model's prediction confidence is systematically lower for a specific user segment?

A Reviewing the serving endpoint's HTTP error rate logs B Monitoring only aggregate mean confidence across all users C Checking model accuracy on the training set weekly D Logging prediction confidence scores disaggregated by user segment and alerting when mean confidence falls below segment-specific thresholds

54 / 60

Question 55 of 60

On the Google Professional ML Engineer 2023 exam, which testing practice MOST effectively validates ML pipeline code changes before merging into the main branch?

A Manual review by the model owner only B Running only linting and type-checking on pipeline code files C Deploying directly to production and monitoring for regressions D Automated CI pipeline running unit tests on each component, integration tests on the full pipeline with a sample dataset, and a canary training run that validates model output quality metrics against minimum thresholds

55 / 60

Question 56 of 60

The Google Professional ML Engineer 2023 exam covers 'speculative decoding' for LLM serving. Which benefit does it provide?

A It reduces training compute by skipping less important training steps B It enables smaller training batch sizes without gradient accumulation C It eliminates the need for GPU accelerators in LLM inference D It accelerates autoregressive LLM inference by having a small draft model generate token candidates that the large model verifies in parallel â€” reducing total generation time

56 / 60

Question 57 of 60

According to the Google Professional ML Engineer 2023 exam, when building a production-grade NLP pipeline that must handle 20+ languages, which approach MOST effectively manages multilingual model quality?

A Training a separate monolingual model for each language B Using English-only training data and relying on translation APIs at inference time C Applying character-level models without any language-specific handling D Fine-tuning a multilingual foundation model (e.g., mBERT, XLM-R) on language-specific labeled data, evaluating performance separately per language, and deploying with language detection routing

57 / 60

Question 58 of 60

The Google Professional ML Engineer 2023 exam asks about designing an ML system for a regulated healthcare application. Which combination of Vertex AI features MOST comprehensively addresses auditability and compliance requirements?

A Deploying to a single region to simplify compliance jurisdiction B Using only Cloud Logging for endpoint access logs C Storing model weights in Cloud Storage with versioned object names D Vertex AI Experiments (full parameter and metric logging), Vertex AI Pipelines (versioned, reproducible training workflows), Vertex AI Model Registry (lineage tracking), Vertex AI Model Monitoring (production drift alerts), and Vertex AI Explainability (per-prediction feature attributions)

58 / 60

Question 59 of 60

According to the Google Professional ML Engineer 2023 exam, which design consideration is MOST important when building an ML system that must generate predictions for regulatory audit at any point in the future?

A Storing only the model's final weights without training metadata B Using the fastest model architecture regardless of interpretability C Resetting model versions after each retraining to reduce storage cost D Maintaining full model lineage (dataset version, pipeline version, model version, serving configuration) and immutable prediction logs with timestamps â€” enabling replay and audit of any prediction

59 / 60

Question 60 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST effectively operationalizes responsible AI principles throughout the ML development lifecycle?

A Conducting a one-time responsible AI review before production deployment B Releasing the model publicly first and addressing responsible AI concerns based on user feedback C Documenting responsible AI principles in a policy document without integrating them into workflows D Embedding responsible AI practices at every lifecycle stage: fairness-aware data collection, bias auditing during evaluation, model card documentation, production monitoring for disparate impact, and structured incident response for harmful outputs

60 / 60

Google ML Engineer Advanced Quiz