Explain the difference between Bayesian and frequentist approaches to statistics. When would you use each in an ML engineering context, and what are the practical tradeoffs?
You mentioned Thompson sampling for bandits. Walk me through how it works and why it's Bayesian — what advantage does it have over epsilon-greedy?
tldr
Frequentist = fixed parameters, probability as long-run frequency, guaranteed error rates. Bayesian = parameters have distributions, probability as belief, formally incorporates prior knowledge. Use frequentist for large-data hypothesis testing and regulated contexts. Use Bayesian for small data, uncertainty quantification, online learning (Thompson sampling), and cold start problems. At ML scale, approximate Bayesian methods (MC Dropout, ensembles) are the pragmatic middle ground.
follow-up
- How would you set a prior for a new ML model's performance when you have no historical data for that specific model but have results from similar models?
- Explain the difference between a confidence interval and a credible interval — and why a product manager would care about the difference.
- When would Bayesian model averaging be preferable to picking a single best model?