You already know: the fundamentals of supervised learning and evaluation (D3). Deep learning and maths answers point to D2 and D4; read those lessons for the full reasoning.
How to use this lesson. Cover the answer, say yours out loud, then compare. A good rapid-fire answer has three parts: the definition, why it matters, and one concrete detail that shows you've used it (a number, a failure mode, a fix). If you can't give the third part, that's the lesson to revisit. Answers are deliberately short; interviewers follow up on whatever you say, so only say what you can defend.
1. Fundamentals (Q1–Q20)
Q1. What is the bias-variance trade-off? Error from wrong assumptions (bias, underfitting) versus error from sensitivity to the particular training sample (variance, overfitting). More complex models lower bias and raise variance; you pick complexity where validation error is lowest.
Q2. How do you detect overfitting? Training score much better than validation score. Learning curves make it visible: a persistent gap that more data narrows.