📝 New paper on multimodal representation learning accepted to TMLR and on generative models for accessibility at CHI 2026

We demonstrate that structural supervision significantly enhances representation learning. By leveraging the inherent redundancy of multimodal and multiview data through joint masked reconstruction and cross-view alignment, we integrate auxiliary signals to improve performance. Checkout our preprint: Structure is Supervision: Multiview Masked Autoencoders for Radiology.

Additionally, we will be presenting our work on adapting generative models for accessibility in collaboration with UNICEF at CHI 2026, checkout the preprint: Steering Generative Models for Accessibility: EasyRead Image Generation.