Regularization (L1/L2) — ML Technical Interview Questions & Answers

Penalizing large weights to reduce overfitting.

Regularization (L1/L2)

  • Simple meaning: Regularisation stops weights from going wild so the model cannot memorise noise. L2 (Ridge) shrinks weights smoothly. L1 (Lasso) can push some weights to exactly 0 (feature drop).
  • Picture: A polynomial that wiggles through every student. Add a penalty for large coefficients. The curve calms down. Test score usually improves if you were overfitting.
  • sklearn: Ridge, Lasso, LogisticRegression(penalty=...). Strength is alpha / C — know which way “more regularisation” goes for the class you use.
  • Trap: regularising when you are underfitting (you will get worse). Trap: mixing up L1 and L2 in a viva — L1 sparsifies, L2 shrinks.

Technical Interview Tip: L2 shrinks. L1 can set weights to 0. Use when train ≫ test.

Code Examples for Regularization (L1/L2)

Regularization (L1/L2) — working example

# Toy train vs test errors (higher = worse)
models = [
    {"name": "underfit", "train": 0.30, "test": 0.32},
    {"name": "good", "train": 0.08, "test": 0.10},
    {"name": "overfit", "train": 0.01, "test": 0.28},
]
for m in models:
    gap = m["test"] - m["train"]
    print(m["name"], "gap=", round(gap, 2))

# L2 idea: penalty += lambda * w**2  (discourages huge weights)

Regularization (L1/L2) — overfit = low train / high test gap; underfit = high error everywhere; regularization limits capacity.

How to Explain Regularization (L1/L2) to an Interviewer

L1 vs L2?

L1 can zero weights L2 shrinks all

When regularise?

Overfit Train much better than test