Download Google Cloud Professional Machine Learning Engineer Ultimate Exam and more Exams Technology in PDF only on Docsity!
Learning Engineer Ultimate Exam
Question 1. Which BigQuery ML model type is most appropriate for predicting a continuous numeric target such as house price? A) Logistic Regression B) K-means Clustering C) Linear Regression D) XGBoost Answer: C Explanation: Linear Regression in BQML is designed for continuous numeric predictions, whereas Logistic Regression is for binary classification, K-means is unsupervised clustering, and XGBoost is a gradient-boosted tree model that can also handle regression but the simplest and most direct choice for a numeric target is Linear Regression. Question 2. In BigQuery ML, which SQL clause is used to split data into training and evaluation sets automatically? A) CREATE MODEL … OPTIONS (model_type='linear_reg') B) SELECT … FROM … WHERE RAND() < 0. C) CREATE MODEL … OPTIONS (split_method='random', eval_fraction=0.2) D) SELECT … FROM ML.EVALUATE Answer: C Explanation: The split_method and eval_fraction options let BQML automatically partition the dataset into training and evaluation subsets. Using RAND() manually is possible but not the built-in split feature. Question 3. When using BigQuery ML’s ARIMA_PLUS model, which parameter controls the seasonality length of the time series?
Learning Engineer Ultimate Exam
A) time_series_timestamp_col B) time_series_data_col C) time_series_id_col D) holiday_region Answer: A Explanation: time_series_timestamp_col supplies the timestamps, enabling ARIMA_PLUS to detect and model seasonal patterns based on the interval between timestamps. The other options relate to data values, grouping, or holiday effects. Question 4. Which of the following statements about Vertex AI Model Garden is FALSE? A) It provides pre-trained large language models such as Gemini. B) Models from the garden can be deployed directly to Vertex AI Endpoints. C) All models in the garden require custom training before deployment. D) You can fine-tune some garden models on your own dataset. Answer: C Explanation: Models in the Model Garden can be used as-is, deployed directly, or fine-tuned; they do not all require custom training first. Hence statement C is false. Question 5. Which Vision API feature is best suited for extracting text from scanned documents? A) LABEL_DETECTION B) FACE_DETECTION C) TEXT_DETECTION
Learning Engineer Ultimate Exam
Explanation: If the required classes are not part of the pre-trained API’s taxonomy, a custom model is necessary. Training epochs, storage size, and region are secondary considerations. Question 8. AutoML Tables automatically performs which of the following steps? A) Manual feature engineering by the user. B) Hyperparameter optimization across multiple model families. C) Deployment to on-premises hardware. D) Real-time data ingestion from Pub/Sub. Answer: B Explanation: AutoML Tables conducts automated feature engineering, model selection, and hyperparameter tuning across several model families (e.g., boosted trees, neural nets). It does not require manual feature work, nor does it handle deployment to on-prem or real-time ingestion directly. Question 9. Which Vertex AI labeling service option is most appropriate for a dataset of 10,000 images that need bounding-box annotations? A) Text classification labeling. B) Image classification labeling. C) Object detection labeling. D) Video classification labeling. Answer: C Explanation: Object detection labeling provides tools for drawing bounding boxes around objects within images, matching the requirement. Text classification and image classification do not capture location information.
Learning Engineer Ultimate Exam
Question 10. What is the primary trade-off when choosing AutoML over custom training? A) AutoML provides higher model accuracy but requires more code. B) AutoML reduces time-to-market but offers less control over model architecture. C) AutoML requires you to manage Kubernetes clusters manually. D) AutoML eliminates the need for any data preprocessing. Answer: B Explanation: AutoML accelerates development by handling many steps automatically, but you sacrifice fine-grained control over architecture, hyperparameters, and custom preprocessing pipelines. Question 11. Which tool can be used to visually identify bias and imbalance in a dataset stored in BigQuery? A) Vertex AI Workbench B) Dataflow C) Facets Overview D) Cloud Scheduler Answer: C Explanation: Facets Overview (or Facets Dive) provides interactive visualizations to detect class imbalance, feature distribution anomalies, and potential bias directly from BigQuery tables. Question 12. To preprocess streaming data at scale before feeding it to a training pipeline, which managed service is most appropriate? A) Cloud Functions B) Cloud Dataflow (Apache Beam)
Learning Engineer Ultimate Exam
Answer: B Explanation: Vertex AI Experiments is specifically designed to log and compare experiment runs, capturing parameters, metrics, and artifacts for reproducibility. Question 15. In a shared VPC environment, which IAM role grants a machine learning engineer read-only access to datasets in a different project? A) roles/bigquery.dataViewer B) roles/owner C) roles/ml.admin D) roles/storage.objectAdmin Answer: A Explanation: roles/bigquery.dataViewer provides read-only access to BigQuery datasets across projects. Owner gives full control, ml.admin is for Vertex AI resources, and storage.objectAdmin is for Cloud Storage. Question 16. Which managed notebook environment offers pre-installed TensorFlow, scikit-learn, and integration with Vertex AI? A) AI Platform Notebooks (legacy) B) Vertex AI Workbench C) Cloud Shell D) Cloud Functions Answer: B
Learning Engineer Ultimate Exam
Explanation: Vertex AI Workbench (formerly AI Platform Notebooks) provides managed JupyterLab notebooks with common ML libraries and direct integration with Vertex AI services. Question 17. When packaging a custom training container, which file defines the runtime dependencies? A) requirements.txt B) Dockerfile C) setup.py D) cloudbuild.yaml Answer: B Explanation: The Dockerfile specifies the base image, OS packages, Python libraries, and entrypoint for the container. requirements.txt may be referenced inside the Dockerfile, but the Dockerfile is the definitive definition. Question 18. Which TensorFlow distribution strategy is best suited for training on a single machine with multiple GPUs? A) MirroredStrategy B) MultiWorkerMirroredStrategy C) ParameterServerStrategy D) CentralStorageStrategy Answer: A Explanation: MirroredStrategy replicates the model on each GPU within a single machine and synchronizes gradients, making it ideal for multi-GPU single-node training.
Learning Engineer Ultimate Exam
C) Monitoring model drift. D) Generating synthetic data. Answer: B Explanation: Vizier is Google’s black-box optimization service integrated with Vertex AI to automatically explore hyperparameter spaces and suggest optimal configurations. Question 22. Quantization of a neural network primarily reduces which aspect of the model? A) Number of layers. B) Training time. C) Model size and inference latency. D) Accuracy on the training set. Answer: C Explanation: Quantization converts weights from 32-bit floating point to lower-precision formats (e.g., INT8), decreasing model size and improving inference speed, often with minimal impact on accuracy. Question 23. Which fine-tuning technique modifies only a small set of additional parameters while keeping the base LLM weights frozen? A) Full fine-tuning B) LoRA (Low-Rank Adaptation) C) Knowledge Distillation D) Weight Pruning Answer: B
Learning Engineer Ultimate Exam
Explanation: LoRA adds low-rank trainable matrices to existing layers, allowing efficient adaptation without altering the original weights, unlike full fine-tuning. Question 24. In Vertex AI Endpoints, which scaling option automatically adjusts the number of node replicas based on request latency? A) Manual scaling B) Autoscaling with traffic percentage C) Autoscaling with CPU utilization target D) Autoscaling with request latency target Answer: D Explanation: Autoscaling can be configured with a latency target, prompting the service to add or remove node replicas to keep prediction latency within the defined threshold. Question 25. For processing 5 TB of historical data to generate predictions in a single batch job, which method is most cost-effective? A) Vertex AI Online Prediction B) Vertex AI Batch Prediction using Cloud Storage as input/output C) Real-time streaming predictions via Pub/Sub D) Deploying a custom endpoint with 100 GPUs Answer: B Explanation: Batch Prediction reads data from Cloud Storage, processes it in parallel, and writes results back, optimizing cost for large, non-real-time workloads. Online prediction would be prohibitively expensive at this scale.
Learning Engineer Ultimate Exam
C) production D) experimental Answer: D Explanation: Vertex AI supports custom aliases such as latest, stable, and production; experimental is not a reserved or built-in alias (though you could create it manually, it is not a predefined type). Question 29. When building a Vertex AI Pipeline, which component type is responsible for executing arbitrary Python code? A) ContainerOp B) CustomJobOp C) NotebookOp D) PythonOp Answer: D Explanation: PythonOp runs Python functions directly within the pipeline, whereas ContainerOp runs a container image, CustomJobOp launches a Vertex AI Custom Job, and NotebookOp runs a notebook. Question 30. Which Google Cloud service can trigger a Vertex AI Pipeline automatically when a new CSV file lands in a Cloud Storage bucket? A) Cloud Scheduler B) Cloud Pub/Sub + Cloud Functions C) Cloud Build D) Cloud Composer Answer: B
Learning Engineer Ultimate Exam
Explanation: A Cloud Storage event can publish a message to Pub/Sub, which a Cloud Function can listen to and then invoke the Vertex AI Pipeline via the SDK. Question 31. In a CI/CD workflow for ML, which step ensures that a new model does not degrade performance compared to the previous version? A) Unit test of preprocessing scripts. B) Integration test of data ingestion. C) Regression test of model metrics on a validation set. D) Linting of Python code. Answer: C Explanation: Regression testing compares key performance metrics (e.g., accuracy, AUC) of the new model against the baseline to detect degradations before promotion. Question 32. Which Vertex AI service records lineage information such as which dataset produced which model artifact? A) Vertex AI Experiments B) Vertex AI Model Monitoring C) Vertex ML Metadata D) Vertex AI Workbench Answer: C Explanation: Vertex ML Metadata captures data, model, and pipeline artifacts along with their relationships, enabling full lineage tracking. Question 33. Feature drift detection in Vertex AI Model Monitoring is triggered when:
Learning Engineer Ultimate Exam
D) Hyperparameter tuning Answer: B Explanation: Safety filters (content moderation) are designed to screen and block disallowed or harmful outputs from generative models. Question 36. Shapley values in Vertex AI Explainable AI are used to: A) Optimize hyperparameters. B) Quantify each feature’s contribution to a specific prediction. C) Detect data drift. D) Encrypt model weights. Answer: B Explanation: Shapley values assign an importance score to each feature for an individual prediction, offering interpretability. Question 37. Which responsible AI toolkit component evaluates model fairness across protected attributes? A) Bias detection module. B) Model compression module. C) Hyperparameter search module. D) Data versioning module. Answer: A Explanation: The bias detection module assesses disparate impact and other fairness metrics across defined protected groups. Question 38. In Vertex AI, what is the purpose of a model alias?
Learning Engineer Ultimate Exam
A) To store model hyperparameters. B) To provide a stable identifier that can point to different model versions over time. C) To encrypt the model binary. D) To automatically scale the model endpoint. Answer: B Explanation: An alias is a human-readable name (e.g., production) that can be reassigned to a newer model version without changing client code. Question 39. Which of the following is a recommended practice when using Vertex AI Feature Store for online serving? A) Store all features as strings to simplify retrieval. B) Enable TTL (time-to-live) on feature rows to expire stale data. C) Write features directly from the online store during batch training. D) Use a single global feature for all entities. Answer: B Explanation: Setting a TTL helps ensure that outdated feature values are automatically removed, preventing stale data from being served. Question 40. When creating a Dataflow pipeline for preprocessing, which runner should you select for production execution? A) DirectRunner B) SparkRunner C) DataflowRunner D) FlinkRunner
Learning Engineer Ultimate Exam
A) TPUs have higher single-core clock speed. B) TPUs provide higher matrix multiplication throughput for large batch sizes. C) TPUs are cheaper per hour in all regions. D) TPUs support Windows OS. Answer: B Explanation: TPUs excel at large-batch matrix multiplications common in transformer training, offering higher throughput than GPUs for such workloads. Clock speed, cost, and OS support are not primary advantages. Question 44. What is the default storage location for model artifacts created by Vertex AI Custom Training jobs? A) Cloud Storage bucket ml-artifacts in the same project. B) Vertex AI Model Registry. C) Cloud SQL instance. D) BigQuery dataset. Answer: B Explanation: After a Custom Training job finishes, the trained model is automatically uploaded to the Vertex AI Model Registry, where it can be versioned and deployed. Question 45. Which of the following is NOT a supported input format for Vertex AI Batch Prediction? A) CSV files in Cloud Storage. B) JSONL files in Cloud Storage. C) Avro files in Cloud Storage. D) Parquet files in BigQuery.
Learning Engineer Ultimate Exam
Answer: C Explanation: Vertex AI Batch Prediction currently supports CSV, JSONL, and BigQuery tables (including Parquet via external tables), but Avro is not a supported direct input format. Question 46. To ensure reproducibility of a training pipeline, which artifact should be version-controlled? A) Docker image tag. B) Random seed only. C) Compute Engine machine type. D) Cloud Logging logs. Answer: A Explanation: Pinning the exact Docker image (including tag or digest) guarantees the same runtime environment across runs, a key factor for reproducibility. Question 47. Which Vertex AI service provides automated data labeling for image classification tasks? A. Vertex AI Data Labeling Service (now part of Vertex AI) B. AutoML Video C. Cloud Data Catalog D. Cloud DLP Answer: A Explanation: The Vertex AI Data Labeling Service enables human-in-the-loop labeling for images, text, and video, integrating directly with AutoML and custom training pipelines.
