# Learning Mechanics

> The mathematical science of neural network training

Learning Mechanics is a research education site focused on the science of deep learning — understanding *why* neural networks work the way they do. We write in-depth posts and tutorials aimed at researchers and practitioners who want to go beyond intuition and engage with the underlying mathematics.

## Posts

- [On neural scaling and the quanta hypothesis](https://learningmechanics.pub/quanta): What is the origin of neural scaling laws? What do they tell us about the structure of data? What are the limits of interpretability?
- [Deep linear networks are a surprisingly useful toy model of weight-space dynamics](https://learningmechanics.pub/deep-linear-nets): Deep linear networks are simple enough to study analytically but rich enough to exhibit key phenomena of neural network training.
- [A visual guide to progressive sharpening and the edge of stability](https://learningmechanics.pub/progressive-sharpening): How neural networks learn in stages, sharpening their representations progressively.
- [The scientific method in two steps](https://learningmechanics.pub/perspectives/scientific-method)
- [Science plays the long game](https://learningmechanics.pub/perspectives/science-is-a-long-game): Fundamental science as playing the long game
- [Towards an atlas of deep learning](https://learningmechanics.pub/perspectives/science-as-mapmaking)

## Other pages

- [Open Questions](https://learningmechanics.pub/openquestions): A curated list of open research questions and broad directions in the science of deep learning.
- [About](https://learningmechanics.pub/about): About Learning Mechanics.

## About

Learning Mechanics is written by researchers at UC Berkeley and collaborators. Posts are peer-reviewed and mathematically rigorous.