Evolution and extension of subspace modularity

Determine how subspace structure and dimensionality evolve across training and across network layers, and establish whether the modularity hierarchy extends naturally to multi-way decompositions involving more than two functions.

Background

The experiments analyze learned decompositions at selected layers and focus primarily on separating two functions at a time. The authors explicitly identify the evolution of subspace organization during training, its variation across layers, and extension from pairwise to multi-function decompositions as unresolved questions.