Beyond SAEs: Consistent Concept Discovery with Sparse Coding
Replaces the linear SAE encoder with per-sample sparse coding to learn dictionaries that stay consistent across training runs.
jinho-choi
Jinho Choi
•
•
1 min read
VisualScratchpad: Grounding Visual Concepts in Large Vision Language Models
Grounding visual concepts in large vision-language models via a attention-based linking mechanism.
hyesu-lim
ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts
Characterizes dataset bias by disentangling visual concepts learned by sparse autoencoders on vision models.
jinho-choi
Sparse Autoencoders Reveal Selective Remapping of Visual Concepts During Adaptation
Sparse autoencoders expose how visual concepts are selectively remapped when vision models adapt to new domains.
hyesu-lim