Answer practical ML foundations questions

In a machine learning interview, you may face a set of applied ML fundamentals questions:

1. The model produces probability outputs. When is calibration required, and what approach would you use to calibrate? How can you assess calibration quality?
2. Feature selection: why is it useful, and which practical techniques do you apply (filter, wrapper, embedded)? How might you carry out feature selection with neural networks?
3. Tokenizers: what does a tokenizer do, what are the common variants (BPE, WordPiece, unigram), and which practical trade-offs are important?
4. Optimizers: give a high-level explanation of Adam (which statistics it tracks and why), and describe situations where it may fail or require tuning.

Give clear, applied answers with concrete examples.

Overview: This question tests a candidate's hands-on command of ML fundamentals, including probability calibration and how to evaluate it, feature selection approaches such as neural-network-based methods, tokenizer types and their NLP trade-offs, and optimizer behavior like the statistics Adam maintains.