Researchers pitch 'learning mechanics' as deep learning's physics

A group of fourteen researchers has put out a position paper making a claim that would have sounded fringe a couple of years ago, that a scientific theory of deep learning is actually emerging and deserves a name. Their pick, in the paper posted to arXiv, is 'learning mechanics.'

The claim is not that we now understand neural networks the way physicists understand a pendulum. It is that five separate strands of work have started to converge, and together they look more like the early skeleton of a real theory than a pile of anecdotes. The authors list them plainly: solvable idealized settings, tractable limits (the infinite width regime being the famous example), simple mathematical laws that capture macroscopic behavior like neural scaling laws, theories that disentangle what hyperparameters actually do, and universal behaviors that show up across systems and settings. The scaling laws point is the one most practitioners already believe. The others are the ones academic theorists have been quietly building.

The physics analogy is doing a lot of the work in the paper's rhetoric. The authors write that 'where mechanistic interpretability aims to be the biology of deep learning, learning mechanics should aspire to be its physics.' That framing is more than branding. It is a claim about the division of labor in a growing field. Interpretability groups pull apart the internals of a specific model the way biologists dissect a specific organism. Learning mechanics would try to describe the forces that shape any model as it trains, closer to classical or statistical mechanics than to anatomy.

The honest caveat is that this is a position paper, not a result. It does not commit to specific empirical predictions that would falsify the program, and the seven desiderata it lays out (fundamental, mathematical, predictive, comprehensive, intuitive, useful, humble) are aspirational, not a scorecard. Take the framing as an invitation to a research community, not proof that the theory has arrived. The paper itself asks readers to be humble about limits, and that hedge is worth keeping.

If it works, the payoff is less exotic than 'we understand neural networks' and more practical. A shared vocabulary and shared desiderata make it easier for labs to hire theorists, for grants to fund coherent programs, and for engineers to tell load-bearing explanations from folklore. Whether the name 'learning mechanics' sticks is less interesting than whether the community that gathers under it actually converges on the same questions.