ConceptGaussian closure
Gaussian closure
The theoretical assumption that, early in learning, the statistical distribution of overlaps can be approximated by a Gaussian distribution, thereby reducing high-dimensional learning dynamics to a small number of collective variables.
Timeline (last 90 days)
Oct 19
The researchers estimate the specialization timescale, i.e., the moment when the Gaussian closure fails and individual learned features begin to strongly correlate with specific target features.
Unverified50%