Katkat
Back

Explain overfitting, regularization, and LLM techniques

AmazonSoftware Engineer · Technical Screen
MediumSoftware EngineerTechnical ScreenMachine Learning

Last updated

This question tests the conceptual understanding of model generalization trade-offs, the mathematical mechanics of regularization, and the practical application of modern LLM architectures.

Part 1: Model Generalization and Regularization

Explain the concept of the bias-variance tradeoff. If a model exhibits high variance (overfitting), how do L1 (Lasso) and L2 (Ridge) regularization mathematically influence the model weights, and in what specific scenario would you prefer one over the other?

Part 2: LLM Techniques

You are tasked with building a domain-specific Q&A bot using a large pre-trained model. Compare and contrast the following approaches in terms of latency, compute cost, and knowledge freshness:

  1. LoRA (Low-Rank Adaptation): When would you choose this over full fine-tuning?
  2. RAG (Retrieval-Augmented Generation): How does this mitigate hallucinations compared to fine-tuning?
  3. Agents: Describe a scenario where an agentic workflow (using tools/APIs) is superior to a zero-shot prompt.

Part 3: Applied Computer Vision

You are designing a computer vision system to detect manufacturing defects on a high-speed conveyor belt.

  • How would you handle a highly imbalanced dataset where 99% of images are 'non-defective'?
  • What metrics would you prioritize (e.g., Precision, Recall, F1, ROC-AUC) and why?
  • If the model performs well in testing but fails in production, what are the most likely 'data drift' or 'environmental' causes?

Hints

Hint 1

Consider how L1 regularization introduces sparsity in weight matrices by penalizing the absolute value of coefficients.

Hint 2

Think about the difference between 'internalizing' knowledge during weights updates (fine-tuning) versus 'referencing' external ground truth via retrieval (RAG).

Hint 3

For the imbalanced CV task, consider techniques like oversampling, cost-sensitive learning, or moving beyond simple accuracy to PR-curves.

Follow-up questions

  • How does batch normalization impact training stability and convergence speed?
  • If your RAG system retrieves irrelevant context, how would you troubleshoot the retrieval pipeline versus the generation pipeline?
  • How would you approach model monitoring for a vision model deployed in a facility where lighting conditions change throughout the day?