Building intuition for optimizers in deep learning
What optimizers actually do, how to choose one, and how to read training by eye. Charlie Maier · SAI-SPARKS · October 2026.
W ← W − η · P(∇L)
Optimizer playground
GD, momentum, Newton, Adagrad, RMSProp, Adam and sign descent on 2D loss surfaces. Presets 1–9 follow the talk. Runs in any browser, offline.
Notebooks
Run on Colab's free CPU; outputs are saved, so they read fine without running.
Slides
The talk deck (PowerPoint), with speaker notes.
Use it with your agent
The repo's AGENTS.md tells a coding agent how to log and plot these diagnostics for your own model.