Explain ML fundamentals (activations, CV, vision, sorting)
Company: Dandy
Role: Machine Learning Engineer
Category: Machine Learning
Difficulty: medium
Interview Round: Technical Screen
You are asked several ML-fundamentals questions. Answer each clearly and concisely, including key assumptions, trade-offs, and what you would do in practice.
1. **Neural networks:** What happens if you replace *all* activation functions in a neural network with the identity function \(y=x\)?
2. **Model evaluation:** Compare **LOOCV (leave-one-out / N-fold CV)** vs **5-fold cross-validation**. How do they differ when estimating **test-time generalization error**? Discuss bias/variance, compute cost, and when you would prefer one.
3. **Computer vision / geometry:** A single photo contains multiple people. Camera focal length and subject distance are unknown. How would you predict each person’s **height**? What **loss function** would you use (and why)?
4. **Neural networks for sorting:** Given an input list of floating-point numbers, how would you design a neural network that outputs the numbers **sorted**? Describe an approach that is trainable end-to-end and how you would define the training objective.
Quick Answer: This question evaluates core machine learning competencies including neural network behavior and activation functions, model evaluation and cross-validation trade-offs, geometric reasoning for computer vision and loss selection, and the design of end-to-end neural architectures for algorithmic tasks such as sorting.