Professional Machine Learning Engineer — Practice Questions and Mock Exam

Sub-word tokenization methods like SentencePiece and WordPiece are standard for modern NLP. TensorFlow Text implements three sub-word tokenizers: BertTokenizer (using WordPiece, as in BERT), WordpieceTokenizer, and SentencepieceTokenizer (used by T5, mT5). Sub-word tokenization breaks words into meaningful units (e.g., 'searchability' → 'search ##ability'), handling out-of-vocabulary words gracefully while keeping vocabulary sizes manageable (30K-100K tokens). WordPiece uses a greedy longest-match-first algorithm; SentencePiece is language-agnostic and handles raw text without pre-tokenization. Word-level tokenization fails on unseen words. Character-level creates excessively long sequences. Rule-based tokenization cannot generalize across diverse languages. Google's Gemini and PaLM models use SentencePiece. Exam tip: modern NLP tokenization = sub-word (SentencePiece/WordPiece/BPE). Ref: www.tensorflow.org/text/guide/subwords_tokenizer

Gemini Enterprise Agent Platform ML Metadata captures ML pipeline metadata as a graph with three entity types: Artifacts (datasets, models, metrics), Executions (pipeline steps), and Contexts (pipeline runs), connected by Events. This creates a queryable lineage graph answering questions like: 'Which dataset was used to train this model?' and 'What hyperparameters produced the most accurate model?' ML Metadata is always on — it automatically records all input/output artifacts from every pipeline run. Experiments on Agent Platform tracks parameters and metrics for experiment comparison but lacks full pipeline-level lineage. Cloud Audit Logs capture API calls for security, not ML artifact relationships. Model Registry manages model versions but doesn't trace back to datasets. Exam tip: 'lineage' or 'provenance' in ML pipelines → Gemini Enterprise Agent Platform ML Metadata. Ref: docs.cloud.google.com/vertex-ai/docs/ml-metadata/introduction

Training-serving skew at the feature level commonly occurs when batch and streaming feature computation pipelines use different code. Feature Store faithfully stores and serves whatever values are written — if the batch pipeline computes a feature using one aggregation or normalization and the streaming pipeline uses a different implementation, the model receives different feature values at serving time than during training. Google's docs note: 'Without a featurestore, you might have different code paths for generating features between training and serving.' The solution is shared computation logic, ideally using Apache Beam which supports both batch and streaming modes. Feature Store doesn't restrict mixed ingestion, and its serving layer works correctly. Overfitting is a model problem, not a feature pipeline issue. Exam tip: different feature values online vs. training → check computation logic parity. Ref: docs.cloud.google.com/vertex-ai/docs/featurestore/overview

BigQuery ML lets you build models using SQL directly where data resides. For tabular classification like churn prediction, BOOSTED_TREE_CLASSIFIER (powered by XGBoost) is effective and requires minimal code: a single CREATE MODEL statement. Data never leaves BigQuery, eliminating pipeline complexity. BigQuery ML supports linear/logistic regression, k-means, ARIMA_PLUS, boosted trees, DNN, and matrix factorization. Exporting to Cloud Storage for custom TensorFlow training adds unnecessary infrastructure overhead. AutoML Tables requires data export. Pre-trained Model Garden models target unstructured data (text, images), not structured tabular data. Exam tip: structured/tabular data already in BigQuery + low-code → BigQuery ML. Ref: docs.cloud.google.com/bigquery/docs/reference/standard-sql/bigqueryml-syntax-create-boosted-tree