Neural language models are really good at explaining held-out data. So when we sample from them, why do they yield dull and degenerate text? Our paper analyzes this behavior using information theory, and corrects for it with a new sampling principle: arxiv.org/abs/2202.00666

[New paper 🚨] You’ve heard of catastrophic forgetting, but have you heard of *fortuitous* forgetting? In this paper, we show how forgetting can be a friend of learning in artificial neural nets. With Ankit Vani, Hugo Larochelle, Aaron Courville (1/6) arxiv.org/abs/2202.00155

Feeling numb. Devastated. Yesterday was Saraswati Puja & today Ma took her blessed one with her. Somehow it feels that even the birds, trees & wind are silent today. Swar Kokila Bharat Ratna #LataMangeshkar ji your divine voice will echo till eternity. Rest in peace. Om Shanti.

Happy to share that a part of my Dual Degree Project work “Semantics-Driven Generative Replay for Few-Shot Class Incremental Learning" (Biplab Banerjee @fabio_cuzzolin) has been accepted ACM Multimedia.

Excited to share that our work “A-STAR: Test-time Attention Segregation and Retention for Text-to-image Synthesis” has been accepted at #ICCV2023 Paper: arxiv.org/abs/2306.14544

Do check out the latest work on improved text-to-image diffusion model by Aishwarya Agarwal and several other esteemed colleagues at Adobe Research India. #ICCV2023

Excited to share that I'll be attending #ICCV2023 in-person! See you at our poster on Wednesday (10:30 AM - 12:30 PM). Looking forward to meeting you all! #ICCV2025 Adobe Research

Excited to be at #ICCV2025 to present two papers: 1) openaccess.thecvf.com/content/ICCV20… 2) openaccess.thecvf.com/content/ICCV20… Also, we are looking for awesome Computer Vision interns to join us at #adoberesearch. Meet Kuldeep Kulkarni, Aishwarya Agarwal or myself for a chat. Merci!

Discover Adobe Research’s new work in #ComputerVision at #ICCV2023 this week! In addition to technical papers, Adobe authors have also contributed to the conference in many other ways. Check out our contributions: adobe.ly/3ZHZESj

Introducing EMILIE: Iterative Multi-granular Image Editing using Diffusion Models. paper Accepted at IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Work led by Joseph KJ Prateksha Udhayanan with Aishwarya Agarwal Balaji Vasan Srinivasan #WACV2024 arxiv.org/pdf/2309.00613…

Introducing MATTE - a Multi-ATTributE Inversion algorithm to constrain image generation of pre-trained stable diffusion models on different attributes (e.g., color, style, layout, semantics) of user-provided reference images. arXiv: arxiv.org/abs/2311.11919

To all the prospective PhDs in/from India, before you sign up for a PhD position in the US, please carefully consider if you would do better (PhDing) in India. I sketched my experience in this blog vihari.github.io/why_india/. Feel free to send me your comments 😃

Unlock your new era of productivity 🔓 with the new AI Assistant (beta) in Adobe Acrobat. Now available on Acrobat desktop and web; coming soon to Reader. ✨ Learn more: adobe.ly/48wgFBK

Poster session from 6:15-8:15 pm tonight at the Exhibit Hall!

Here's a link to a lecture on interpretability that I gave at Dawn Song and UCB course last spring, which covers a some of the ways we're thinking about AI interpretability in our lab. youtube.com/watch?v=8QPVKp… Thanks Dan and Dawn for the opportunity to talk!

🚀 Just launched a new repo compiling key papers on Explainable CV! 🔍 I've added short summary for each paper, with a major focus on concept-based explanations—I'm expanding to include other directions too! Contributions are welcome! Check it out here: github.com/aisagarw/aweso…

Today, I reached 100 citations on Google Scholar! It might seem a bit trivial to celebrate this, but looking at where I started just a couple years back, it really does feel like a meaningful milestone. scholar.google.com/citations?user…

Disappointed with your ICLR paper being rejected? Ten years ago today, Sergey and I finished training some of the first end-to-end neutral nets for robot control 🤖 We submitted the paper to RSS on January 23, 2015. It was rejected for being "incremental" and "unlikely to have

Discover how Adobe Researchers joined forces with the Adobe Express team to create On-brand Recolor—a new feature that uses AI and tailored algorithms to thoughtfully recolor graphics with your brand’s color palette in mind! adobe.ly/4jx9ks9

Excited to share that our work “TIDE: Training Locally Interpretable Domain Generalization Models Enables Test-time Correction” has been accepted at #CVPR2025 🎉 Paper: arxiv.org/abs/2411.16788