Loss-Invariant Projections as Passive Probes of Learned Representations

Akshay Chandrasekhar, Pavlo Melnyk

Learned feature representations in neural networks often contain structure beyond that directly used by the final task output. We study this structure using $\textit{passive probes…