A suite of Vision Sparse Autoencoders — LessWrongCLIP-Scope? Inspired by Gemma-Scope We trained 8 Sparse Autoencoders each on 1.2 billion tokens on different layers of a Vision Transformer. These (a…https://www.lesswrong.com/posts/wrznNDMRmbQABAEMH/a-suite-of-vision-sparse-autoencodersNSFW Featuresmats_sae_training_for_ViTsHugoFry • Updated 2024 Oct 21 2:55Towards Multimodal Interpretability: Learning Sparse Interpretable Features in Vision Transformers — LessWrongExecutive Summary In this post I present my results from training a Sparse Autoencoder (SAE) on a CLIP Vision Transformer (ViT) using the ImageNet-1k…https://www.lesswrong.com/posts/bCtbuWraqYTDtuARg/towards-multimodal-interpretability-learning-sparse-2web appThis is a streamlit app to allow users to explore features learnt by a Sparse AutoEncoder (SAE) t...https://sae-explorer.streamlit.app/Task Vectors are Cross-ModalTask Vectors are Cross-ModalTask representations in VLMs are consistent across modality (text, image) and specification (example, instruction).https://task-vectors-are-cross-modal.github.io/
Seonglae ChoSeonglae ChoSeonglae ChoSeonglae Cho
