Explain overfitting, regularization, and LLM techniques
Quick Overview
This question evaluates understanding of model generalization (overfitting vs underfitting), regularization methods (L1 vs L2), modern LLM techniques (LoRA, RAG, agents), and end-to-end computer vision project skills including dataset construction, evaluation, failure modes, and deployment, testing competencies across ML theory and engineering.
Amazon
Feb 12, 2026, 12:00 AM
Software Engineer
Technical Screen
Machine Learning
3
0
Loading...
You’re in an ML interview. Answer the following conceptual questions clearly and concisely, using examples where helpful:
1) Model fit
- What is overfitting vs underfitting ?
- For each, list common symptoms you would see in training/validation curves.
- Give 3–5 practical ways to mitigate each problem.
2) Regularization
-
Compare
L1
vs
L2
regularization:
- objective/penalty form
- effect on weights (sparsity vs shrinkage)
- when you would prefer one over the other
- interaction with correlated features
3) LLM-related topics
Explain the purpose, core idea, and major trade-offs for:
- LoRA (low-rank adaptation) for fine-tuning
- RAG (retrieval-augmented generation)
- Agents (tool-use / planning loops)
For each, describe:
- what problem it solves
- what data it needs
- what can go wrong (failure modes)
- how you would evaluate it in production
4) Project deep dive (CV example)
Pick one computer-vision project you’ve worked on (e.g., classification/detection/segmentation) and be prepared to explain:
- problem statement and business goal
- dataset construction/labeling and leakage risks
- model choice and baseline
- training details (augmentation, loss, class imbalance, hyperparameters)
- evaluation metrics and thresholding
- key errors you found and how you fixed them
- how you would deploy/monitor it (latency, drift, feedback loop)
Solution
Show