Aman Mehra

I'm a ML Scientist at Tesla Autopilot where I build models that power FSD. I am most excited about real world robotic problems tackling safety and efficacy in the 99% performance band. To this end, I am interested in topics spanning generalization, generative modeling and sequential decision making.

Previously, I spent two years in beautiful Pittsburgh as a MS Robotics student at Carnegie Mellon University, advised by Prof. Jeff Schneider. In addition, I had the pleasure of collaborating with Prof. Aditi Raghunathan and Prof. Zico Kolter.

Preceeding this was a wonderful year, working on image generation and perspective geometry, as a research fellow at Microsoft Research, under the mentorship of the amazing Saikat Guha.

Even earlier, I was a CS undergrad at IIIT Delhi working with Prof. Mayank Vatsa, Prof. Richa Singh and Prof. Sanjit Kaul

CV / Google Scholar / Github

Research

	Auxiliary Start States for Accelerating Online Reinforcement Learning Aman Mehra, Alexandre Capone, Jeff Schneider ICML ARLET Workshop, 2024 PDF Smart strategies for sampling of reset states to accelerate online reinforcement learning. Particularly useful in sparse reward, hard exploration problems.
	Predicting the Performance of Foundation Models via Agreement-on-the-line Aman Mehra, Rahul Saxena, Taeyoun Kim*, Christina Baek, J Zico Kolter, Aditi Raghunathan NeurIPS, 2024 NeurIPS R0-FoMo Workshop , 2023 NeurIPS DistShift Workshop , 2023 Arxiv / Poster Understanding when can accuracy of finetuned foundation models be reliably estimated out-of-distribution with the phenomemon of Agreement-on-the-Line.
	Motion Magnified 3D Residual-in-Dense Network for Deepfake Detection Aman Mehra, Akshay Agarwal, Mayank Vatsa, Richa Singh IEEE Transactions on Biometrics, Behavior, and Identity Science , 2022 Best Student Paper Award Link / PDF / Bibtex Amplifying temporal inconsistencies to detect deepfake videos with confounding JPEG compression artifacts.
	Detection of Digital Manipulation in Facial Images (Student Abstract) Aman Mehra, Akshay Agarwal, Mayank Vatsa, Richa Singh AAAI (Student Abstract) , 2021 Link / Bibtex Detecting deepfake videos through 3D CNNs.

Projects

	Sparse Voxel Grids for Accelerated Inverse Rendering Aman Mehra Report, 2023 PDF Accelerated inverse rendering from posed multi-view images via sparse voxel grids.
	Lottery Tickets through the lens of Random Matrix Theory Aman Mehra Report, 2021 PDF A study to probe whether analysis from random matrix theory can characterize winning lotteries in deep networks.
	Refresh Stale Transitions in Replay Buffer Aman Mehra Report, 2021 PDF/ Code Improving sample efficiency in off-policy RL by refreshing stale transitions in replay buffer.
	CUDA accelerated RRT Aman Mehra, Paras Sharma, Savit Gupta Report, 2021 PDF / Code An implementation of the RRT algorithm accelerated using Quad Trees and CUDA