Vatsal Baherwani

I am a visiting researcher at the Center for Human-Compatible AI at UC Berkeley. I am also starting my PhD at NYU in Fall 2025. I previously graduated with a bachelor's degree from the University of Maryland.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github  /  LinkedIn  /  Blog

profile photo

Research
Video Diffusion Models Encode Motion in Early Timesteps
Vatsal Baherwani, Yixuan Ren, Abhinav Shrivastava
Under Review

We show that motion information is independently learned in early timesteps of the diffusion process, prior to the materialization of spatial attributes. We use this insight to present a simple and efficient method for targeted video motion customization.

Dense Backpropagation Improves Routing for Sparsely-Gated Mixture-of-Experts
Ashwinee Panda*, Vatsal Baherwani*, Zain Sarwar, Benjamin Therien, Supriyo Chakraborty, Tom Goldstein
NeurIPS 2024 OPT, ENLSP, Neural Compression Workshops

We approximate the dense backward pass of a sparse mixture-of-experts model, leading to improved training stability and performance with negligible overhead.

Racial and Gender Stereotypes Encoded Into CLIP Representations
Vatsal Baherwani, Joseph Vincent
ICLR 2024 Tiny Paper

We demonstrate that image and text embeddings in CLIP exhibit strong biases corresponding to prevalent racial and gender stereotypes.


Thanks to Jon Barron for his website template.