A scientific theory of deep learning is emerging, slowly but surely.

Understanding deep learning will be the intellectual challenge of the early 21st century, much like understanding quantum mechanics in the early 20th. Progress is being made: pieces of a scientific theory of deep learning are starting to be uncovered and fit together. It’s a slow process, and one aided by coordination among different groups, so we’ve made this website as a hub to organize and share progress.

Why a science of deep learning? As it matures, this emerging science will become practically impactful in the training and usage of large models, and also (we anticipate) a central tool for AI safety and alignment. Plus it’s really fascinating: it has deep connections to our own learning. See the learning mechanics perspective paper for a fuller argument and a description of the emerging science.

If you want to write an article for Learning Mechanics, please reach out to the editors.


Editors

Daniel Kunin
Daniel KuninUC Berkeley
Florentin Guth
Florentin GuthNYU and the Flatiron Institute
Jamie Simon
Jamie SimonUC Berkeley and Imbue

Team

Enric Boix-Adserà
Enric Boix-AdseràUniversity of Pennsylvania
Jeremy Cohen
Jeremy CohenFlatiron Institute
Nikhil Ghosh
Nikhil GhoshFlatiron Institute
Florentin Guth
Florentin GuthNYU and the Flatiron Institute
Arthur Jacot
Arthur JacotNew York University
Mason Kamb
Mason KambStanford
Dhruva Karkada
Dhruva KarkadaUC Berkeley
Daniel Kunin
Daniel KuninUC Berkeley
Berkan Ottlik
Berkan OttlikUniversity of Pennsylvania
Jamie Simon
Jamie SimonUC Berkeley and Imbue
Joey Turnbull
Joey TurnbullUC Berkeley