JO
joschu.net
Summary
joschu.net covering Reinforcement Learning, KL Divergence, Monte Carlo Estimation, and Variance Reduction. Notable guests include John Schulman. Episodes span from Jan 2020 to Mar 2020.
2episodes
16total claims
12topics covered
2 episodes
Approximating KL Divergence
›Mar 7, 2020
The standard Monte Carlo estimator for KL divergence, `log(q(x)/p(x))`, is unbiased but suffers from high variance, making it unreliable for practical applications like reinforcement learning diagn...
KL DivergenceMonte Carlo EstimationVariance ReductionBias-Variance Tradeoff+9 more
An Opinionated Guide to ML Research
›Jan 24, 2020 · WITH John Schulman
The author, John Schulman, advocates for a 'goal-driven' research approach, where researchers define a novel AI capability and work towards it, rather than an 'idea-driven' approach of making incre...
Machine Learning ResearchResearch MethodologyGoal-Driven ResearchIdea-Driven Research+11 more