Explain core ML fundamentals and tradeoffs
Quick Overview
This question evaluates core machine learning fundamentals including bias–variance tradeoffs, overfitting, class imbalance handling, loss function selection, optimization algorithms, and high-level neural network architecture choices, testing competencies in model evaluation, training dynamics, regularization, and robustness within the Machine Learning domain. It is commonly asked because employers need to assess conceptual understanding alongside practical application for production-oriented tasks like recommendation, ranking, and classification, specifically the ability to reason about trade-offs, diagnostics, and techniques that impact model performance and deployment.
ML Fundamentals Interview Prompt
Answer the following ML fundamentals questions clearly and with practical examples:
-
Bias vs. variance
- What are bias and variance?
- How do you diagnose high bias vs high variance from train/validation curves?
- What actions reduce bias vs reduce variance?
-
Overfitting
- Why does overfitting happen?
- List common mitigations for linear models and for neural networks.
-
Imbalanced data
- Why can accuracy be misleading?
- What metrics are better?
- What approaches can you use at the data level, algorithm level, and thresholding level?
-
Loss functions
(especially for neural networks)
- When would you use MSE vs cross-entropy vs focal loss?
- What is label smoothing and why might it help?
-
Optimization
- Compare SGD, Momentum, Adam.
- What are learning-rate schedules and why do they matter?
- What problems do vanishing/exploding gradients cause and how do you address them?
-
Neural network architectures (high level)
- When would you prefer CNNs, RNNs, Transformers?
- What are common regularization techniques (dropout, weight decay, batch norm) and how do they work?
Assume a product ML setting (recommendation/ranking/classification).