Explain Core ML Training, Evaluation, and Transformer Concepts

# Explain Core ML Training, Evaluation, and Transformer Concepts Answer the following machine-learning fundamentals questions. Use concrete examples and connect each technique to the failure mode it addresses. ### Clarifying Questions to Ask - Is the evaluation setting binary classification, multiclass classification, or regression? - Is class imbalance present in training data, deployment traffic, or both? - Should the optimizer discussion compare update rules mathematically or focus on practical selection? - Is the Transformer explanation expected to cover an encoder, an autoregressive decoder, or the shared building blocks? ### Part 1 — Reduce Overfitting How would you diagnose and reduce overfitting? Distinguish interventions that add data, constrain the model, alter training, or improve validation. #### What This Part Should Cover - A train-versus-validation generalization gap rather than training loss alone. - Regularization, early stopping, simpler capacity, augmentation, and leakage-safe validation. - Selection based on held-out behavior instead of repeated tuning on a final test set. ### Part 2 — Handle Class Imbalance How would you train and evaluate a classifier when the positive class is rare? #### What This Part Should Cover - Stratified or time-appropriate splits and metrics aligned with false-positive and false-negative costs. - Class weighting, resampling, threshold selection, and probability calibration. - Avoidance of duplicated validation examples or synthetic leakage. ### Part 3 — Explain Optimizer Fundamentals Compare plain stochastic gradient descent, momentum, and an adaptive optimizer such as Adam. Explain the roles of learning rate, gradient scale, state, and convergence diagnostics. #### What This Part Should Cover - Parameter updates based on gradients and the effect of noisy mini-batches. - Momentum accumulation and adaptive per-parameter scaling. - Learning-rate schedules, weight decay, and signs of divergence or stalled training. ### Part 4 — Define Evaluation Metrics Define precision, recall, F1, ROC-AUC, PR-AUC, and log loss for classification. Explain when accuracy is misleading and how you would choose a threshold-dependent metric. #### What This Part Should Cover - Correct confusion-matrix formulas for precision and recall. - The harmonic mean in F1 and the ranking interpretations of the two AUC families. - Log loss as a probability-quality metric and metric choice tied to operating costs. ### Part 5 — Introduce the Transformer Give a concise explanation of a Transformer block and how token information moves through it. Include position information, self-attention, feed-forward layers, residual paths, normalization, masking, and the sequence-length cost of dense attention. #### What This Part Should Cover - Query, key, and value projections and normalized attention weights. - Multi-head representations plus position information. - Residual and normalization structure, causal masking where applicable, and quadratic dense-attention cost. ### What a Strong Answer Covers - Correct definitions joined to practical decisions and failure modes. - A clear boundary between changing training distribution and changing decision thresholds. - Evaluation procedures that prevent leakage and reflect deployment prevalence. - Transformer mechanics without claiming that attention alone supplies token order. - Trade-offs rather than a single technique presented as universally best. ### Follow-up Questions - How would you detect that an apparent imbalance improvement only changed calibration? - Why can oversampling and class weighting have similar gradient effects but different operational costs? - What changes in self-attention during autoregressive inference with a key-value cache?