TODO: provide reference, possibly a podcast?
An interesting claim about the arrangement of data points into clusters that can make things easier or harder to perform classification tasks. I found this interesting because the representations created by DCCs are tightly clustered and segregated by default.
Description
Describe neural classification task by regions called manifolds or their manifold representation
The manifold is defined as the Perceptual Manifold
Small manifolds (tightly clustered data points) make easy classification.
Aligned manifolds also make classification easier.
Geometrically aligned manifolds make the task of “disentanglement” easier as well, finding the similar traits between different classes.
Two aspects of Recognition: Disentanglement and Classification
Similar to a jar of coins, where the coins are manifolds. Disorganized coins have gaps between them, but tightly organized and aligned coins have little space between them and are orderly and more coins can fit