High-performing neural network models of visual cortex benefit from high latent dimensionality
Fig 3
Relationship between effective dimensionality and encoding performance.
a: The encoding performance achieved by a model scaled with its effective dimensionality. This trend also held within different kinds of model training paradigms (supervised, self-supervised, untrained). Each point in the plot was obtained from one layer from one DNN, resulting in a total of 568 models. Encoding performance is for the monkey IT brain region, which had the strongest relationship with ED among regions we considered. b: Even when conditioning on a particular DNN layer, controlling for both depth and ambient dimensionality (i.e., number of neurons), effective dimensionality and encoding performance continued to strongly correlate. The plot shows the distribution of these correlations (Pearson r) across all unique layers in our analyses. c,d: Similar results were obtained for human fMRI data.