Encyclopedia · 176 concepts

Math & Optimization · intermediate · concept 46 of 176

Information Theory: Entropy & KL Divergence

Shannon's mathematics of surprise. Entropy measures how unpredictable a distribution is, cross-entropy is the loss nearly every classifier and LLM trains on, and KL divergence measures how far one distribution drifts from another, which is how RLHF keeps a tuned model close to its base.

Key terms

EntropyCross-entropy lossKL divergenceBitsPerplexity

Where you meet it in the real world

The loss function of LLMs, compression, perplexity benchmarks, the KL penalty in RLHF

Videos

Entropy (for data science) Clearly Explained!!!

StatQuest with Josh Starmer · YouTube

Guides and articles

Visual Information Theory -- colah's blog

Chris Olah (Anthropic co-founder)