<ul><li>Sparse autoencoders (SAEs) have helped address the understanding of vision and language processing mechanisms.</li><li>SAEs trained on CLIP's vision transformer reveal distinct sparsity patterns across layers and token types.</li><li>Metrics are introduced to quantify the steerability of SAE features, with 10-15% of neurons and features being steerable.</li><li>Targeted suppression of SAE features improves performance on vision disentanglement tasks and defense against typographic attacks.</li></ul>

Steering CLIP's vision transformer with sparse autoencoders

Discover more