Developed guided diffusion and flow matching algorithms for linear bandit prior sampling, as well as low-rank approximations to reduce per-round time complexity while preserving regret bounds.
Open-source code available on
Github.
A from-scratch implementation of both traditional auto-regressive and discrete diffusion generative transformer-based language models.
Trained on openwebtext utilizing DGX system.
Open-source code available on
Github.
A from-scratch implementation of denoising diffusion and flow matching image generative models (CNN U-Net and latent vision transformer).
Trained on Stanford Cars, CelebA, ImageNet utilizing DGX system.
Open-source code available on
Github.