Steering CLIP's vision transformer with sparse autoencoders
원문 초록
While vision models are highly capable, their internal mechanisms remain poorly understood -- a challenge which sparse autoencoders (SAEs) have helped address in language, but which remains underexplored in vision. We address this gap by training SAEs on CLIP's vision transformer and uncover key differences between vision and language processing, including distinct sparsity patterns for SAEs trained across layers and token types. We then provide the first systematic analysis on the steerability of CLIP's vision transformer by introducing metrics to quantify how precisely SAE features can be steered to affect the model's output. We find that 10-15\% of neurons and features are steerable, with SAEs providing thousands more steerable features than the base model. Through targeted suppression of SAE features, we then demonstrate improved performance on three vision disentanglement tasks (CelebA, Waterbirds, and typographic attacks), finding optimal disentanglement in middle model layers, and achieving state-of-the-art performance on defense against typographic attacks.
함께 살펴볼 논문
공통 태그를 근거로 연결했습니다. 실제 인용 관계나 논문의 우수성을 뜻하지 않습니다.
Adversarial Disentanglement by Backpropagation with Physics-Informed Variational Autoencoder
Dual-View Disentangled Multi-Intent Learning for Enhanced Collaborative Filtering
Instruction-Tuned Video-Audio Models Elucidate Functional Specialization in the Brain
Revisiting Cross-Modal Knowledge Distillation: A Disentanglement Approach for RGBD Semantic Segmentation
Causal-LLaVA: Causal Disentanglement for Mitigating Hallucination in Multimodal Large Language Models
Interpretability Illusions with Sparse Autoencoders: Evaluating Robustness of Concept Representations
토론
이 논문과 연결된 포럼 글이 아직 없습니다.