One of my current interests is the theoretical and applied study of diffusion models. I am currently working on improved techniques for score estimation to improve the quality (sample quality and likelihood) of diffusion models and implicit models. I am also interested in new methods for one-step generative modeling using novel diffusion distillation techniques based on maximum mean discrepancy (MMD). I am also interested and have applied diffusion models for solving inverse problems, such as those based on structured additive noise forward processes. I am also working on efforts that involve the distributed training of large scale transformer architectures across multiple H100 GPUs for signal source separation.

At the same time, I have also been doing research on new representation learning techniques that can leverage spectral decomposition techniques for applications to progressive image decoding and variable bitrate compression. Another application of such representation learning techniques is data attribution in neural network architectures such as transformer-based LLMs or CNNs. During my Masters, I developed a new architecture for lossless and lossy image compression by designing a novel algorithm for compression with model-code separated compression architectures (i.e., source-agnostic encoding and source augmentable decoding). During my time at UIUC, I worked on developing new video frame interpolation technology by leveraging the Local All-Pass (LAP) optical flow estimation algorithm.