2.1 Supervised Learning Basics
Interview-ready field notes
The 20-second answer
Supervised learning learns a mapping from labeled inputs to targets: regression predicts a number, classification predicts a class or class probability.
Core ideas to retain
- Train data fits weights, validation data selects hyperparameters, and test data is the final unbiased check.
- Regression commonly uses MAE or MSE; classification commonly uses cross-entropy.
- Underfitting means high train and validation error; overfitting means low train error but worse validation error.
- Bias is error from an overly simple assumption; variance is sensitivity to the sampled training data.
Interview / OA rule
If both train and validation loss are bad, increase useful capacity or improve features. If only validation is bad, regularize, add data, or simplify the model.
One good written resource
Google — Machine Learning Crash Course — read this after the video when you want a clearer mental model, not more pages of notes.
Most asked
Interview questions to practise aloud
Each answer is the level of detail expected for a strong fundamentals round.
Regression vs classification?
Regression predicts continuous values; classification predicts labels or probabilities over labels. Their output activations and losses usually differ.
What is the bias-variance trade-off?
More flexible models can lower bias but may raise variance. Good generalization balances both using data, validation, and regularization.
Why keep a test set?
Repeated choices based on a test score leak information from that test set. Use validation for choices and test once at the end.