A coalition of researchers led by Daniel Kunin released a paper arguing that a scientific theory of deep learning is coalescing. The authors define this framework as a set of laws describing training dynamics, hidden representations, weight distributions, and model performance. They label the emerging perspective learning mechanics.

The manuscript groups five research strands that collectively push the theory forward. First, idealized solvable models grant intuition about real‑world learning curves. Second, tractable limits reveal core phenomena such as double descent. Third, simple macroscopic laws capture observable metrics like loss decay. Fourth, hyperparameter analyses isolate their effect, and fifth, universal behaviors across architectures hint at deeper regularities.

All five strands share a focus on training dynamics and coarse‑grained statistics, emphasizing falsifiable quantitative predictions. By treating deep learning as a mechanical process, the authors contrast their view with statistical and information‑theoretic approaches, proposing a symbiotic link with mechanistic interpretability research. The paper also rebuts common skepticism about the feasibility or relevance of a formal theory.

Concluding, the authors outline open problems in learning mechanics, from extending solvable regimes to integrating interpretability insights. They provide a repository of introductory materials and invite newcomers to explore the outlined directions, positioning the work as a roadmap for future theoretical advances in neural network science.