My research focusses on Out-Of-Distribution generalization---enabling AI to perform well in situations not encountered in the training data. In my PhD I have worked on:- AI Algorithms: Developed algorithms to improve generalization capabilities of AI models under distribution shifts in 3D viewpoints, scene lighting, and object materials, among others.- Machine Learning Theory: Developed theoretical underpinnings of OOD generalization to identify attributes of robust machine learning representations.- Comparing Humans and AI agents: Humans serve as an upper bound for AI. We benchmarked how well modern vision models and LLMs compare to humans under distribution shifts. [8, 9]- Generalization in Macaque Brains: Investigated generalization in-vivo through Electro-physiological recordings from Macaque Brains to understand when and how brains generalize under distribution shifts.Full list of publications: https://scholar.google.com/citations?user=QY5OAIMAAAAJ

Teaching fellow for Professor Gabriel Kreiman's class on Biological and Artificial Intelligence (Neuro 240):- Awarded a Teaching Distinction Award by the Derek Bok Center for Teaching for my role.- The class was a semester-long project based class, and I was single handedly in-charge of helping over 40 students with their projects. I was awarded a Teaching Award by the Derek Bok Center for

Investigating if generative AI models can create out-of-distribution lighting: - Implemented a procedural graphics pipeline to create photorealistic cities under changing lighting.- Investigated lighting variations generated by Generative AI models (GANs) trained on the above dataset to evaluate utility of GenAI for Adobe’s photo-relighting efforts.

Worked in collaboration with MIT on building large-scale, controlled datasets to evaluate AI models.- Computer Graphics: Implemented a procedural graphics pipeline to create photorealistic cities with fine-grained control over scene parameters including object viewpoints, light source distribution and intensity, and scene layout, resulting a novel dataset to study behaviour Computer Vision models.- Generalization in AI: Investigated generalization behaviour of visual recognition models using above mentioned dataset, leading to a publication

Worked on building ChatBots with a Persona.- Developed an algorithm for non-parametric reservoir sampling—given a seed data distribution, our algorithm searches and picks samples from the internet to increase data size with minimal changes to the data distribution.- Created custom conversational agents for Microsoft Xbox. These agents used the above algorithm to build their own training datasets by sampling relevant lines similar to a given seed.

Worked on Machine Learning models for targeted genome editign:- Conceptualized and implemented an algorithm for designing targeted molecular scissors (zinc proteins) for cleaving DNA at desired target locations.- Using a mixture of synthetic and experimental data generated using molecular docking simulations, trained an ensemble of neural networks to predict DNA-Protein interactions.This work led to two publications (available here: https://scholar.google.com/citations?user=QY5OAIMAAAAJ)