ActSub decomposes activations via SVD of the classifier head into decisive and insignificant subspaces, using cosine similarity on the insignificant part for far-OOD and shaped energy on the decisive part for near-OOD, achieving SOTA on standard benchmarks.
Deep residual learning for image recognition
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Activation Subspaces for Out-of-Distribution Detection
ActSub decomposes activations via SVD of the classifier head into decisive and insignificant subspaces, using cosine similarity on the insignificant part for far-OOD and shaped energy on the decisive part for near-OOD, achieving SOTA on standard benchmarks.