We visualized the features of 16 SAEs trained on CLIP in collaboration between Fraunhofer HHI and Mila - Institut québécois d'IA! Search thousands of interpretable CLIP features in our vision atlas, with autointerp labels, & scores like clarity and polysemanticity. Some fun features in thread:

Link:

Some features are creative and unexpected, like this "adoptee" feature of humans and dogs with their babies.

This "fictional animal" captures specifically captures photoshopped animals and could be useful for fraud detection.

We also have some features that feel almost poetic, like this medley of open lipstick-ed mouths and tiger mouths.

While we mostly filtered the dataset for explicit images, there is a good collection of bombs, guns, and other weapons. You can find them if you look for them--- relevant for screening models for dangerous content and safety issues.

We will release all the SAEs open source soon as part of the Prisma library, along with our whitepaper! Meanwhile, you can browse CLIP's feature atlas here: Do screenshot and post/send me unusual features!

Thank you to my wonderful collaborators @praneet_suresh_ & @LorenzHufe, and the incredible team at @FraunhoferHHI, including Jim Berend, Maximilian Dreyer, Tobias Labarta, Johanna Vielhaben, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek, among others, for making this happen!

QLA Castle Seminar is now open for reservation meatheads! I have perfected this methodology using decades of experience in creating generational wealth as fast as humanly possible for myself and my mentees. #QLA #TheTrillionDollarMan

@FraunhoferHHI @Mila_Quebec Great job

@FraunhoferHHI @Mila_Quebec 👏👏👏

@FraunhoferHHI @Mila_Quebec This looks amazing! Are you going to release the auto-interp pipeline you all used? I wanted to try applying it to some other models.