Finally, we also performed zero-shot object localization by applying COCOA to image & text representations from a CLIP model. (6/6)

By visualizing important features that make an augmentation-transformed image similar to the original image (as opposed to random images), we showcased the utility of COCOA for understanding SimCLR representations under data augmentation. (5/n)

We showed that COCOA performed well for explaining unsupervised models. For example, removing pixels with high COCOA scores makes the downstream predicted probability decrease fast, whereas insertion makes the predicted probability increase fast. (4/n)

At the core is the *contrastive corpus similarity* between a sample and the contrastive direction toward the corpus and away from the foil in the learned representation space. This similarity can be paired with various feature attribution methods to generate COCOA scores. (3/n).

The tl;dr is: We proposed COntrastive COrpus Attribution (COCOA), which identifies input features that are contrastively important for making a sample’s learned representation closer to a *corpus* set and further from a *foil* set. (2/n).

We have an upcoming paper at ICLR 2023 on a new feature attribution method for explaining representations learned by unsupervised models!.This was joint work with the fantastic @HughChen18 @ChanwooKim_ and my advisor @suinleelab. (1/n)

RT @efweinberger: Excited to share early results from a fantastic collaboration with @chrislin97! Over the past few months, we've been work….