2 ms·
Look at VLM mechanistic interpretability papers vs just pca on JEPA trained weights. JEPA gives you interpretability for free. I have not personally inspected
by vatsachak 4mo ago
Look at VLM mechanistic interpretability papers vs just pca on JEPA trained weights.
JEPA gives you interpretability for free.
I have not personally inspected them and my view is maybe a more exaggerated/dramatic claim of those working in the JEPA sphere
- hssa45 4mo agoSounds interesting, any links?
- vatsachak 4mo agoJEPA in chess leads to interpretable chess boards: https://arxiv.org/abs/2606.11860 https://arxiv.org/abs/2606.11860 JEPA in image classification leads to interpretable image latents https://arxiv.org/abs/2508.10104 https://arxiv.org/abs/2508.10104 Easy intro to JEPA, demonstrating that interpretability is as easy as running a PCA on latents https://youtu.be/kYkIdXwW2AE?is=CfCBcy1jLt-FfI2E https://youtu.be/kYkIdXwW2AE?is=CfCBcy1jLt-FfI2E