Jan 2021 – Jul 2022 · 1 yr 6 mos · Bengaluru, Karnataka, India · On-site

• Researched learning-based approaches for black-box optimization problems based on the idea of imitating the non-differential black-box using a suitable differentiable neural network.
• Worked on a research project based on PIM (Process-in-Memory) which aimed to speed up neural network training and inference using hardware optimizations. Included building and maintaining a low-level C++ library for the new hardware component.
• Built user-friendly tools for layer-wise profiling and gradient statistics visualization for pytorch-based models.

May 2019 – Jul 2019 · 2 mos · Bangalore

• Tested multiple bottom-up, top-down 2D and 3D human pose estimation algorithms.
• Found specific cases where the accuracy of these algorithms drops significantly and suggested possible workarounds.

May 2018 – Dec 2018 · 7 mos

• Worked on controlling a quadrotor using neural networks which are trained using reinforcement learning techniques (imitation learning and Deep Deterministic Policy Gradient).