Ian Covert

Making this class with Su-In, Hugh and Chris was one of the most fun things I did in grad school. We covered a ton of material, definitely check out all the slides we made courses.cs.washington.edu/courses/csep59… I'm excited to see how the course evolves in the next couple years!

📣 📣 📣 Our new paper investigates the question of how many images 🖼️ of a concept are required by a diffusion model 🤖 to imitate it. This question is critical for understanding and mitigating the copyright and privacy infringements of these models! arxiv.org/abs/2410.15002

Very excited to introduce locality alignment, an efficient post-training algorithm to improve your ViTs + VLMs, essentially for free🚀 Local align = new self-supervised objective ensuring that encoder captures fine-grained spatial info. No new data needed. Here's the idea 1/3

How to perform dynamic feature selection without assumptions about the data distribution or fitting generative models? We develop a learning approach to estimate the conditional mutual information in a discriminative fashion for selecting features. arxiv.org/pdf/2306.03301…

This was work done with @HughChen18 @scottlundberg and of course our advisor @suinleelab NMI version: rdcu.be/dcICX arXiv version: arxiv.org/abs/2207.07605

Large models are tough because you may not be able to query the model thousands of times to get attributions (e.g., KernelSHAP). This is something we've tackled in a couple other papers FastSHAP (ICLR'22): arxiv.org/abs/2107.07436 ViT Shapley (ICLR'23): arxiv.org/abs/2206.05282

In our recent Nature MI paper, we looked at the surprising number of algorithms that estimate Shapley values (whose computation scales exponentially with the number of players). There are a lot, we counted at least 24 papers on this topic! Paper: rdcu.be/dcICX

My amazing PhD student @ianccovert will present our work on ViT Shapley at #ICLR2023 soon -- Mon 1 May 11 CAT!

We have an upcoming paper at ICLR 2023 on a new feature attribution method for explaining representations learned by unsupervised models! arxiv.org/abs/2210.00107 This was joint work with the fantastic @HughChen18 @ChanwooKim_ and my advisor @suinleelab. (1/n)

Check out this course on #XAI by @suinleelab & @ianccovert. Very practical and nicely curated. Also points to some great papers on the topic. The course covers a broad set of principles and techniques. Slides are available here: buff.ly/3Dasssm #ResponsibleAI

This was joint work with the fantastic @ChanwooKim_ and @suinleelab from @uwcse/@uw_wail. And big shoutout to collaborators @SudarshanMukund @neiljethani + Rajesh from the paper this work builds on arxiv.org/abs/2107.07436 (9/9)

Our experiments used three image datasets and models as big as ViT-Large (arXiv needs to be updated), but there's still plenty of room to scale this up. My guess is that 1) it gets better with more data, 2) it can help learn better representations than the original task (8/n)

The question we’re trying to answer is *which patches influence the prediction.* And Shapley values are a surprisingly simple approach: they're like leave-one-out, but the effect of removing a patch is averaged across all sets of preceding patches (3/n)

TL;DR: we fine-tune a ViT to directly predict its Shapley values, without using a dataset of ground truth examples. And it works quite well (2/n)

If you want to know what your ViT pays attention to...you might not want to use attention values! Shapley values can do this better, and now they can even do it efficiently. Check out our new paper (ICLR spotlight) arxiv.org/abs/2206.05282 🧵⬇️