Improving the Student-Teacher Approach for Semi-Supervised Semantic Segmentation

Arun Ramachandran, Jeremiah Coholich, Harsh Maheshwari, Simar Kareer

We propose and evaluate two variants to the mean teacher self-distillation technique: sliding-window teacher evaluation and the ensembling of two mean-teachers with different update rates. The latter modification improved results over the baseline mean-teacher.

fasttensor: A C++ high-performance computing library

I wrote a C++ library for high-performance computing of basic arithmetic operations on tensors. It uses SIMD instructions on CPU and CUDA on GPU to accelerate computation, and can use multiple GPUs. It uses expression templates and kernel fusion to avoid unnecessary memory allocations and copies. It achieves over 1.5x speedup over normal eager evaluation.

Chess AI: A self-learning chess artificial intelligence

I made a self-learning chess AI with reinforcement learning and game tree searching. I used MTD-f, an optimized version of minimax, to speed up game tree searching by 1.5x. The AI ended up successfully learning positional advantages of pieces in different phases of the game.

Code Character: An online AI contest

Code Character is a yearly contest hosted by my undergraduate institution, where players write their own AI to control troops in a real-time strategy game. I designed the game for the 2017 and 2018 editions and led a team of programmers to make a web app to automate submissions and a cross-platform C++ engine for the game. In 2018, we launched the contest and had over 1300 participants from 74 different countries.