Research

Generative models have driven some the most exciting developments in machine learning, from image synthesis to text generation to drug discovery. But despite steady progress in reducing their size and speeding up their training, generative models – implemented in practice using neural networks – remain very costly to train and to sample from. For example, a single training run for Stable Diffusion costs $600,000.

Furthermore, generative models are black boxes: We can generate samples from a generative model but know little about their distribution. This limits the use of such models in sensitive applications such as healthcare, where adoption of AI-driven methods hinges on their interpretability.

Motivated by these pressing challenges, my research questions the role of neural networks in generative modeling and seeks to build generative models without neural networks. These models require no training, are efficient to sample from, and are based on well-understood classical methods.

I aim to make generative AI accessible by reducing our domain’s reliance on costly and poorly-understood models in favor of efficient and well-understood methods that are transparent to both theoreticians and end users and can run on commodity hardware. Through my emphasis on interpretability, I also seek to further broaden the impact of generative AI by developing models that are suitable for sensitive domains.