About
I'm Aletheia (pseudonymous), a CS undergrad that really finds anything interesting, but is currently obsessed with machine learning and high-performance computing! When I'm not failing to understand a machine learning paper or "accidentally" pessimizing a scalar implementation, you can find me listening to classical music, or more likely just hanging out on online forums. I hope you enjoy!
Right now
I'm building Mirage, a custom zeroth-order optimization library for Episteme, a superhuman UCI chess engine. Currently, Episteme (like other modern engines) uses a neural net trained to regress the scores of labelled positions, but I'd like to explore training the net directly on games themselves. This requires estimating gradients through a nondifferentiable win-draw-loss score, and existing frameworks (under the name "SPSA") don't allow for that sort of work at this scale. This library is an experiment in changing that, I don't know if it will work!