A new loss term aligns the sparse activation patterns of samples within each class, producing class-consistent and more interpretable latent dimensions in VAEs.
Disentangling disentanglement in variational autoencoders
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Enhancing Interpretability of Sparse Latent Representations with Class Information
A new loss term aligns the sparse activation patterns of samples within each class, producing class-consistent and more interpretable latent dimensions in VAEs.