optimization

Definition

The mathematics and practice of adjusting parameters to minimize a loss — in deep learning, almost always stochastic gradient descent and its variants. Learning rate schedules, momentum, weight decay, and gradient clipping are where much of the practical difficulty lives, and a bad optimizer configuration can hide a good architecture. Distinguish this from engineering optimization, meaning making inference faster.

Explainers using this term

• Paper Explained: Retrieval-Augmented Skill Optimization via Cross-Harness Adaptation — Borrow Someone Else's Skill, Rewrite It for Your Own HarnessRetrieval-Augmented Skill Optimization via Cross-Harness Adaptation
• Paper Explained: RewardVerse — A Video Reward Model That Writes the Rubric Before It Scores (RGPO)RewardVerse: Rubric-Guided Policy Optimization for Video Reward Modeling
• Paper Explained: Bellman Policy Optimization — Policy Mirror Descent for LLMs, Without a Value ModelBellman Policy Optimization

In the textbook

• Making It Follow Instructions — Instruction Tuning and Preference Optimization