I am a post-doctoral researcher in the EI department. I focus on Video generative models and investigate opportunities of causal representations in this context. Apart from that, I am investigating Autoformalizaton, a technique that leverages the vast natural language understanding of LLMs to represent mathematical problems precisely using formal languages where reasoning can be cast as a search problem. I think a program that can come up with a new formal system or at least extend an existing formal system as needed during the runtime can overcome the lack of representation power and therefore applicability of existing formal systems.

Prior to this, I was an IMPRS-IS scholar and was supervised by Michael Black. I focused on Generative modelling. More specifically, the conditional versions of them. Conditional generative models are a means to learn one-to-many mappings. Although such models, especially GANs, do a good job at density estimation of high dimensional data (images), they are far from perfect. These imperfections induce errors that manifest as spurious correlations (among many others) in the generated data. Although this may not be noticeable readily, it can render such models completely unusable for many applications. I worked on addressing this issue.