I'm a PhD student at Northeastern University advised by Prof. David Bau. I'm interested in enumerating and understanding knowledge inside generative image and language models. I also closely work with Prof. Antonio Torralba from MIT. I interned at Adobe Research.
I previously worked at Indian Space Research Organization as an Applied Scientist, where I worked on advancing image sensing capabilites towards better disaster management and monitoring systems using neural networks.
I believe that, as humans, we are only scratching the surface of how we interact with powerful generative models. My research explores how we can go beyond these limits by uncovering novel concepts and designing intuitive ways for people to use these models. The goal is not just to use models, but to use their potential to push the boundaries of human creativity and capability.
Distillated diffusion models are super fast, but lack diversity in output samples. We ask - why? Distilled models have the concept representations required for base model's diversity, but don't use them. Through theoretical analysis and casual experiments we narrow this down to - the first timestep of diffusion generation!
SliderSpace automatically decomposes diffusion models' visual capabilities into controllable, human-understandable directions from a single text prompt. This framework enables users to explore and discover novel concepts encoded inside any diffusion model
We show that language models can be used as classifiers and propose an unlearning methods where a language model self-critics its knowledge and guides itself to unlearn. This method is essential to retain the fluent text generation capabilities after unlearning.
Concept Sliders are light-weight adaptors that can control specific attributes in a diffusion model's outputs. Training these sliders are very simple - just provide the text prompts of the concept (e.g. "winter weather", "old age", "abstract art"). They can be composed and continuously controlled!
UCE employs a fast closed-form solution for editing text-to-image diffusion models to address bias, copyright, and offensive content simultaneously without retraining. The method enables scalable and concurrent edits by modifying cross attention weights.
This work presents a method for removing specific visual concepts from text-to-image diffusion models while preserving their overall generative capabilities.
Advicing
I've had the oppurtunity to work with and advice some of the most amazing students and folks!
Traditionally erasure is evaluated externally by analyzing generated images. We release a nuanced and rigorous suite of evaluation techniques including incontext, training-free, and dynamic tracing probes to investigate if concepts are "really" erased. Answer: Not really!
This work presents a new paradigm - "do diffusion models need millions of artworks in training data to actually learn art?". We release a new diffusion model trained completely on art-free data. This model can still mimic art by seeing less than 5 art images.
Knowledge editing in LLMs have unintended ripple effects on nearby concepts. Measuring this requires building a custom dataset - challenging!. We propose RippleBench - our RAG-LLM pipeline can automatically generate structured datasets given a concept ("biology").