Explain BatchNorm, optimizers, and L1/L2
Company: Snapchat
Role: Machine Learning Engineer
Category: Machine Learning
Difficulty: medium
Interview Round: Onsite
## Prompt
Answer the following ML fundamentals questions:
1. **Batch Normalization (BatchNorm):**
- What trainable parameters does BatchNorm have?
- What statistics are used during training vs inference?
- Why does BatchNorm help optimization, and what are common pitfalls?
2. **Optimizers:** Compare SGD, Momentum, RMSProp, and Adam.
- What problem does each address?
- When might plain SGD generalize better than Adam?
3. **Regularization:** Compare **L1** and **L2** regularization.
- How do they affect weights and sparsity?
- How do they relate to MAP estimation (priors) if you know that framing?
Quick Answer: This question evaluates a candidate's understanding of core machine learning fundamentals—Batch Normalization, optimizer behaviors (SGD, Momentum, RMSProp, Adam), and regularization methods (L1 vs L2)—and the candidate's ability to reason about training dynamics, parameter effects, and inference versus training distinctions in the Machine Learning domain. It is commonly asked in technical interviews because these topics reveal comprehension of optimization dynamics, generalization and sparsity trade-offs, and implementation implications, testing both conceptual understanding and practical application.