Google ML Engineer Interm - Free Practice Quiz

Question 1 of 60

A Google Professional ML Engineer 2023 exam candidate is designing a pipeline for a text classification task. Which approach BEST handles vocabulary terms unseen during training?

A Using subword tokenization (e.g., BPE or WordPiece) to represent unknown words as known subword units B Removing all OOV tokens before inference C Using a fixed vocabulary dictionary and ignoring unknown terms D Retraining the model for each new vocabulary term

1 / 60

Question 2 of 60

On the Google Professional ML Engineer 2023 exam, which technique is used to prevent a neural network from co-adapting neurons and reduce overfitting during training?

A Dropout B Batch normalization C Weight decay D Data augmentation

2 / 60

Question 3 of 60

On the Google Professional ML Engineer 2023 exam, which architecture is MOST appropriate for classifying sequences of text with variable length?

A Transformer-based model (e.g., BERT fine-tuned on the classification task) B Random Forest C Feedforward MLP with fixed input size D Convolutional Neural Network (CNN)

3 / 60

Question 4 of 60

On the Google Professional ML Engineer 2023 exam, which technique MOST effectively prevents training-serving skew when a model relies on real-time features?

A Using a consistent feature serving infrastructure (Vertex AI Feature Store) that serves the same features at training and prediction time B Disabling real-time features and using only static features C Using a single monolithic code path for both training and serving D Recomputing all features from raw data at serving time with separate Python scripts

4 / 60

Question 5 of 60

According to the Google Professional ML Engineer 2023 exam, which Vertex AI capability enables automatically retraining a model when new data becomes available without manual intervention?

A Vertex AI Pipelines with a scheduled trigger or Pub/Sub-driven continuous training pipeline B Vertex AI Workbench running manual notebook cells C Vertex AI Vizier hyperparameter search D Vertex AI Feature Store snapshot export

5 / 60

Question 6 of 60

The Google Professional ML Engineer 2023 exam describes 'streaming ML inference' as MOST appropriate when:

A Predictions must be generated immediately in response to real-time event streams (e.g., fraud detection on transactions) B Results can be computed offline overnight C Predictions are needed once per day for batch reporting D Training data is unavailable for retraining

6 / 60

Question 7 of 60

According to the Google Professional ML Engineer 2023 exam, which technique is used to systematically search for the best model architecture through automated experimentation?

A Neural Architecture Search (NAS) B Random weight initialization C Feature importance analysis D Manual expert-designed architecture selection

7 / 60

Question 8 of 60

The Google Professional ML Engineer 2023 exam addresses 'batch normalization' as a technique that:

A Normalizes the activations of each layer across a mini-batch, accelerating training and reducing sensitivity to weight initialization B Normalizes the input dataset before training begins C Applies regularization to each batch of training data D Divides a dataset into mini-batches for training

8 / 60

Question 9 of 60

The Google Professional ML Engineer 2023 exam asks about 'data versioning' in ML workflows. Which tool is MOST commonly used for versioning large datasets on Google Cloud?

A DVC (Data Version Control) or Vertex AI Dataset versioning in conjunction with Cloud Storage B Git with files committed directly to the repository C Cloud SQL change logs D BigQuery audit logs

9 / 60

Question 10 of 60

According to the Google Professional ML Engineer 2023 exam, which approach BEST detects when a deployed model's predictions have become systematically biased for a demographic group?

A Implementing slice-based performance evaluation that disaggregates metrics by sensitive attributes in monitoring B Reviewing the model's training loss curve C Checking the overall aggregate accuracy metric only D Retraining the model with more epochs

10 / 60

Question 11 of 60

The Google Professional ML Engineer 2023 exam identifies 'catastrophic forgetting' as a challenge in:

A Continual learning, where a model trained sequentially on new tasks forgets previously learned tasks B Deploying a model to a new cloud region C Compressing a model for edge deployment D Training a model from scratch on a new dataset

11 / 60

Question 12 of 60

According to the Google Professional ML Engineer 2023 exam, which approach is MOST effective for building a time-series anomaly detection model on streaming sensor data in Google Cloud?

A Deploying an LSTM or autoencoder model on Vertex AI Online Prediction, consuming Pub/Sub sensor events via Cloud Dataflow for real-time scoring B Using Cloud SQL triggers to flag values exceeding hardcoded thresholds C Storing all sensor readings in BigQuery and running weekly batch anomaly detection D Serverless ML models deployed via Cloud Functions for each sensor individually

12 / 60

Question 13 of 60

On the Google Professional ML Engineer 2023 exam, which technique BEST addresses the vanishing gradient problem in very deep recurrent networks?

A Using LSTM (Long Short-Term Memory) or GRU cells which include gating mechanisms to control gradient flow B Increasing the learning rate significantly C Removing all dropout layers from the network D Using sigmoid activation in all layers

13 / 60

Question 14 of 60

The Google Professional ML Engineer 2023 exam identifies which preprocessing transformation as MOST appropriate for normalizing a skewed numerical feature with a long right tail?

A Log transformation (e.g., log(x+1)) to reduce skewness B Standardization (z-score normalization) C One-hot encoding D Target encoding

14 / 60

Question 15 of 60

On the Google Professional ML Engineer 2023 exam, which Vertex AI component is MOST analogous to a CI/CD system for ML models?

A Vertex AI Pipelines integrated with Cloud Build for automated training, evaluation, and deployment B Vertex AI Feature Store C Vertex AI Model Monitoring D Vertex AI Workbench

15 / 60

Question 16 of 60

The Google Professional ML Engineer 2023 exam identifies 'class activation mapping' (CAM) as a technique PRIMARILY used to:

A Balance class distributions during training B Visualize which regions of an input image most influenced a CNN's classification decision C Convert classification models into regression models D Generate synthetic training images for underrepresented classes

16 / 60

Question 17 of 60

According to the Google Professional ML Engineer 2023 exam, which validation strategy MOST reliably estimates model performance when training data is limited?

A A single 80/20 train/test split B K-fold cross-validation, averaging performance across all folds C Training on 100% of data and evaluating on the same data D Using only 10% of data for training to maximize validation set size

17 / 60

Question 18 of 60

On the Google Professional ML Engineer 2023 exam, which Vertex AI feature allows ML engineers to share reusable, versioned pipeline component definitions across teams?

A Vertex AI Experiments B Vertex AI Pipeline component registry / Artifact Registry for pipeline components C Vertex AI Feature Store D Vertex AI Model Registry

18 / 60

Question 19 of 60

According to the Google Professional ML Engineer 2023 exam, which type of neural network layer is MOST effective for extracting local spatial patterns from image data?

A Dense (fully connected) layer B Convolutional layer (CNN) C Embedding layer D Recurrent layer

19 / 60

Question 20 of 60

On the Google Professional ML Engineer 2023 exam, which approach is MOST appropriate when a customer requires ML model predictions to be explainable for regulatory compliance?

A Using a black-box deep learning model for highest accuracy B Enabling Vertex AI Explainability (SHAP-based or Integrated Gradients) and logging per-prediction feature attributions C Using a linear model or decision tree exclusively since deep learning cannot be explained D Deploying the model without explanation capabilities and providing accuracy metrics

20 / 60

Question 21 of 60

The Google Professional ML Engineer 2023 exam covers 'multi-task learning' as an approach that:

A Distributes training across multiple TPUs B Trains a single model simultaneously on multiple related tasks, sharing representations to improve generalization across all tasks C Trains separate models for each task and averages their predictions D Uses multiple datasets from different domains without task-specific output heads

21 / 60

Question 22 of 60

According to the Google Professional ML Engineer 2023 exam, which approach MOST effectively handles missing values in a training dataset for a tabular regression model?

A Deleting all rows with any missing value before training B Imputing missing values using mean/median imputation for numerical features and mode for categorical, or using model-based imputation (e.g., KNN imputer) C Filling all missing values with zero D Converting missing values to a fixed string 'UNKNOWN'

22 / 60

Question 23 of 60

The Google Professional ML Engineer 2023 exam describes 'model pruning' as a technique that:

A Adds new layers to a model to increase capacity B Removes redundant or low-magnitude weights from a trained model to reduce its size and inference cost C Increases regularization by adding dropout layers D Resets all model weights to random values

23 / 60

Question 24 of 60

According to the Google Professional ML Engineer 2023 exam, what is the PRIMARY advantage of using TFX (TensorFlow Extended) for production ML pipelines?

A TFX is exclusively for NLP models B TFX provides a standardized set of production-grade pipeline components (data validation, transform, training, evaluation, serving) that enforce ML best practices C TFX only runs on-premises D TFX replaces the need for Cloud Storage

24 / 60

Question 25 of 60

On the Google Professional ML Engineer 2023 exam, which approach MOST effectively monitors for feature drift in a Vertex AI deployed model?

A Reviewing the model's prediction output manually each week B Configuring Vertex AI Model Monitoring to compare incoming feature distributions against the training baseline and alert on statistical deviations C Redeploying the model weekly regardless of performance D Checking the endpoint's request latency logs

25 / 60

Question 26 of 60

The Google Professional ML Engineer 2023 exam covers which Google Cloud service for serving large-scale low-latency feature lookups during online ML inference?

A Cloud Spanner B Vertex AI Feature Store (online serving) backed by Bigtable C Cloud Storage D Cloud Dataflow

26 / 60

Question 27 of 60

On the Google Professional ML Engineer 2023 exam, which loss function is MOST appropriate for training a generative adversarial network (GAN)?

A Categorical Cross-Entropy B Adversarial loss â€” the generator minimizes and the discriminator maximizes a value function representing the probability of distinguishing real from fake samples C Huber Loss D Mean Squared Error

27 / 60

Question 28 of 60

The Google Professional ML Engineer 2023 exam covers 'semi-supervised learning' as MOST appropriate when:

A You have large amounts of labeled data and no unlabeled data B A large amount of unlabeled data is available but labeled data is limited and expensive to obtain C All training data must be manually labeled before any model training begins D The target variable is continuous and all data is labeled

28 / 60

Question 29 of 60

According to the Google Professional ML Engineer 2023 exam, which technique is MOST effective for reducing inference latency for a large Transformer model deployed on Vertex AI?

A Increasing batch size at inference time B Model distillation â€” training a smaller student model to replicate the teacher model's predictions C Retraining the model with a higher learning rate D Removing the model's attention layers

29 / 60

Question 30 of 60

On the Google Professional ML Engineer 2023 exam, which hyperparameter tuning strategy is MOST sample-efficient for continuous search spaces?

A Grid search across predefined values B Bayesian optimization using a surrogate model to guide search toward promising regions C Random search with uniform sampling D Exhaustive search of all combinations

30 / 60

Question 31 of 60

According to the Google Professional ML Engineer 2023 exam, when should a Google Cloud ML Engineer prefer Vertex AI AutoML over custom training?

A When the problem requires a custom neural architecture not available in AutoML B When training on more than 10 billion samples C When inference must be served at sub-millisecond latency with a custom serving container D When labeled data is limited, the use case matches AutoML's supported task types, and rapid prototyping is prioritized over full control

31 / 60

Question 32 of 60

On the Google Professional ML Engineer 2023 exam, which pattern is MOST appropriate for serving a computationally expensive ML model to a high-traffic mobile app?

A Each mobile app client directly calls the Vertex AI endpoint for every inference B The model is embedded in the mobile app binary and updated with each app release C All predictions are generated in a nightly batch job and cached for next-day lookup D A lightweight distilled or quantized model is deployed on-device, with fallback to a Vertex AI endpoint for complex cases

32 / 60

Question 33 of 60

The Google Professional ML Engineer 2023 exam addresses 'data augmentation' for tabular data. Which approach is MOST commonly used?

A Converting all numerical columns to categorical bins B Pixel rotation and color jitter applied to tabular rows C Randomly shuffling column headers D SMOTE (Synthetic Minority Over-sampling Technique) for minority class oversampling or Gaussian noise injection into numerical features

33 / 60

Question 34 of 60

According to the Google Professional ML Engineer 2023 exam, which Vertex AI capability enables ML practitioners to interactively label training data and build training datasets?

A Vertex AI Vizier B Vertex AI Workbench C Vertex AI Model Monitoring D Vertex AI Data Labeling

34 / 60

Question 35 of 60

The Google Professional ML Engineer 2023 exam identifies which Cloud Logging and Monitoring integration is MOST important for operational ML model health?

A Logging raw training data to Cloud Logging for auditability B Logging hyperparameter values to Pub/Sub C Exporting model weights to Cloud Logging for recovery D Using Cloud Monitoring to track CPU/memory on the model endpoint and set alerts for latency and error rate anomalies

35 / 60

Question 36 of 60

According to the Google Professional ML Engineer 2023 exam, which approach to ML experiment tracking MOST supports governance and reproducibility requirements?

A Logging metrics in a shared spreadsheet updated manually by the team B Storing experiment results only in developer notebooks C Using only Cloud Logging for experiment metadata D Systematically logging all experiment parameters, metrics, dataset versions, and model artifacts in Vertex AI Experiments with lineage tracking

36 / 60

Question 37 of 60

On the Google Professional ML Engineer 2023 exam, which Vertex AI serving feature allows deploying multiple model versions simultaneously and routing traffic between them for experimentation?

A Vertex AI Vizier parallel trials B Vertex AI Feature Store versioning C Vertex AI Model Monitoring drift detection D Vertex AI Endpoint traffic splitting

37 / 60

Question 38 of 60

The Google Professional ML Engineer 2023 exam identifies 'ensemble methods' as techniques that improve model accuracy by:

A Using a single powerful model architecture trained on a very large dataset B Removing all underperforming neurons during training C Regularizing model weights to reduce complexity D Combining predictions from multiple diverse models to reduce variance and bias

38 / 60

Question 39 of 60

On the Google Professional ML Engineer 2023 exam, 'stratified sampling' when creating train/validation splits ensures that:

A Training data is sorted alphabetically before splitting B All numerical features are normalized before splitting C Validation set is always 10% of total data regardless of class distribution D The proportion of each class is preserved in both train and validation sets, avoiding representation bias

39 / 60

Question 40 of 60

The Google Professional ML Engineer 2023 exam describes 'cold start' in recommendation systems as:

A A model that runs too slowly due to insufficient hardware B Training a model without GPU acceleration C A model performance issue caused by training data containing future leakage D The inability to provide accurate recommendations for new users or new items with no historical interaction data

40 / 60

Question 41 of 60

According to the Google Professional ML Engineer 2023 exam, which approach MOST effectively reduces training cost for large models on Google Cloud?

A Disabling distributed training to reduce overhead B Always using the most powerful on-demand VM type C Reducing training dataset size by 90% D Using preemptible (Spot) VMs or TPU preemptible pods for training jobs that can tolerate interruptions

41 / 60

Question 42 of 60

On the Google Professional ML Engineer 2023 exam, which approach to model evaluation is MOST appropriate for a time-series forecasting model?

A K-fold cross-validation with random fold assignment B Stratified random sampling by date C Leave-one-out cross-validation D Walk-forward validation (expanding window or sliding window evaluation on chronological test folds)

42 / 60

Question 43 of 60

According to the Google Professional ML Engineer 2023 exam, which approach BEST scales a Vertex AI Online Prediction endpoint to handle sudden traffic spikes?

A Pre-configuring a fixed number of replicas and never scaling B Manually increasing replicas when traffic alerts fire C Using a single high-memory machine for all predictions D Enabling autoscaling on the Vertex AI endpoint with min/max replica configuration

43 / 60

Question 44 of 60

On the Google Professional ML Engineer 2023 exam, which technique is used to interpret how a Transformer model decides which input tokens to focus on for a prediction?

A Confusion matrix analysis B LIME explanations C Activation function selection D Attention weight visualization

44 / 60

Question 45 of 60

The Google Professional ML Engineer 2023 exam identifies 'feature importance' analysis as MOST useful for:

A Calculating the model's precision and recall on the test set B Measuring the total number of parameters in the model C Visualizing the model's architecture as a graph D Identifying which input features most significantly influence model predictions, enabling pruning of irrelevant features and improving interpretability

45 / 60

Question 46 of 60

According to the Google Professional ML Engineer 2023 exam, which technique is used in NLP to convert words or subwords into dense numerical vectors that capture semantic relationships?

A TF-IDF sparse vectors B One-hot encoding of vocabulary C Word embeddings or sentence embeddings (e.g., Word2Vec, BERT embeddings) D N-gram bag-of-words representation

46 / 60

Question 47 of 60

The Google Professional ML Engineer 2023 exam describes 'shadow mode deployment' as a strategy where:

A The new model is deployed to a separate region and evaluated on synthetic traffic B The old model is immediately replaced without a phased rollout C Both old and new models receive the same production traffic but only the old model's predictions are served to users, while the new model's predictions are logged for comparison D A feature store snapshot is deployed alongside the model for shadow evaluation

47 / 60

Question 48 of 60

According to the Google Professional ML Engineer 2023 exam, which approach MOST effectively handles concept drift in a fraud detection model deployed in production?

A Deploying the same model indefinitely and relying on rule-based fallbacks B Increasing the fraud detection threshold manually when more fraud is detected C Periodically retraining the model on fresh data using Vertex AI Pipelines continuous training triggered by drift alerts D Adding more features to the original training dataset without retraining

48 / 60

Question 49 of 60

On the Google Professional ML Engineer 2023 exam, which approach BEST ensures reproducibility of a Vertex AI training job?

A Relying on Vertex AI to automatically record all training parameters B Using the latest available base container image each time C Pinning container image versions, fixing random seeds, versioning datasets in Vertex AI Datasets, and tracking all experiment parameters in Vertex AI Experiments D Training models only in Cloud Shell notebooks

49 / 60

Question 50 of 60

The Google Professional ML Engineer 2023 exam describes which MLOps concept as treating the entire ML pipeline (data ingestion, training, evaluation, deployment) as a versioned, auditable software artifact?

A Model versioning only B Feature drift monitoring C ML pipeline as code with full CI/CD integration D A/B testing of individual model layers

50 / 60

Question 51 of 60

On the Google Professional ML Engineer 2023 exam, which technique is used to generate explanations for black-box ML model predictions by approximating the model locally with an interpretable model?

A Shapley values (SHAP) B Attention visualization C LIME (Local Interpretable Model-agnostic Explanations) D Confusion matrix analysis

51 / 60

Question 52 of 60

The Google Professional ML Engineer 2023 exam identifies 'data catalog' tools as important for ML governance because they:

A Replace the need for feature engineering B Automatically train models from cataloged datasets C Provide a searchable inventory of datasets with metadata (schema, lineage, sensitivity) enabling ML teams to discover and reuse high-quality data D Store trained model weights for version control

52 / 60

Question 53 of 60

According to the Google Professional ML Engineer 2023 exam, which loss function is MOST appropriate for a ranking model (learning to rank results by relevance)?

A Binary Cross-Entropy B Mean Squared Error C Pairwise or listwise ranking loss functions (e.g., LambdaRank, ListNet) D Categorical Cross-Entropy

53 / 60

Question 54 of 60

On the Google Professional ML Engineer 2023 exam, which technique MOST effectively reduces overfitting in a gradient boosted tree model?

A Increasing the maximum tree depth to maximum B Removing regularization parameters entirely C Limiting tree depth, reducing learning rate, and using subsampling (row and feature sampling per tree) D Training on only 10% of available data

54 / 60

Question 55 of 60

According to the Google Professional ML Engineer 2023 exam, which Vertex AI tool allows scheduling and monitoring ML training pipelines without writing custom infrastructure code?

A Cloud Composer (Apache Airflow) B Cloud Scheduler with Cloud Functions C Vertex AI Pipelines with a schedule trigger D Cloud Build triggers

55 / 60

Question 56 of 60

On the Google Professional ML Engineer 2023 exam, which serving optimization technique is MOST effective for reducing memory usage of a deployed neural network without significant accuracy loss?

A Increasing the number of model replicas B Retraining the model with a higher learning rate C Post-training quantization (converting float32 weights to int8 or float16) D Removing the validation dataset from training

56 / 60

Question 57 of 60

The Google Professional ML Engineer 2023 exam covers 'model registry' as a component that:

A Transforms raw data into features for training B Monitors model latency in production C Stores versioned, trained model artifacts with metadata (metrics, lineage, deployment status) enabling lifecycle management D Schedules periodic batch prediction jobs

57 / 60

Question 58 of 60

A Vertex AI pipeline produces a trained model artifact, but the ML team cannot determine which dataset version or preprocessing parameters were used to produce it. Which Vertex AI capability directly addresses this traceability requirement?

A Vertex AI Model Registry â€” tracks model versions with performance metrics and deployment history for governance B Vertex AI Feature Store â€” stores and serves feature values with point-in-time consistency for training and online serving C Vertex AI Metadata â€” records artifact lineage, execution parameters, and relationships between datasets, models, and evaluation metrics across all pipeline runs for full reproducibility D Vertex AI Experiments â€” records hyperparameter configurations and evaluation metrics across training runs for comparison and selection

58 / 60

Question 59 of 60

The Google Professional ML Engineer 2023 exam identifies 'curriculum learning' as a training strategy where:

A The model is trained only on the hardest examples from the start B Training data is randomly sampled without any ordering strategy C Training data is presented in order from simpler to more complex examples, mimicking human learning progression D All training examples are presented with equal frequency throughout training

59 / 60

Question 60 of 60

According to the Google Professional ML Engineer 2023 exam, which strategy MOST effectively mitigates bias in a training dataset used for a hiring recommendation model?

A Using only the most recent 6 months of hiring data to avoid historical patterns B Removing the model's output layer and relying on human reviewers only C Auditing training data for protected attribute correlations, applying fairness constraints, and evaluating model performance disaggregated by demographic groups D Training the model on synthetic data generated by a GAN

60 / 60

Google ML Engineer Intermediate Quiz