A new study claims to have found the hidden math behind how neural networks tell things apart.
Researchers analyzed internal representations from top-performing vision, audio, and language models and found that variability within a class is not random. Instead, it correlates strongly with the class's own centroid and with a handful of rival class centroids. From that pattern they built a mean-field theory that predicts classification accuracy using only a sparse set of these centroid coordinates, rather than the full high-dimensional representation. The theory held up across different architectures and data types, and the same handful of geometric measurements that predict accuracy also improve as models scale up.
If a small set of centroid distances can predict how well a model classifies, that is a far simpler explanation than the usual hand-wave that bigger networks just learn better features. It also ties classification behavior to sparse-feature methods like sparse autoencoders, tools the interpretability field already relies on to make black-box models legible.
The math explains correlation, not causation, and grand unifying theories of deep learning have a mixed track record of holding up under scrutiny.