Selected coding and research projects.
Fractal Dimensions and Neural Networks (2021)
It is well established that neural networks can approximate Lipschitz functions. However, real-life data often resembles fractals rather than smooth functions.
Led by a combinatorics professor, my peers and I are trying to prove universal approximation theorem for neural networks to fractals of suitable dimensions, and train neural networks to effectively learn fractal functions.
Video Restoration With Deep Prior (2021)
I aim to restore videos using neural network without a clean dataset. Previously, Ulyanov et. al. demonstrated that randomly initiated deep convolutional neural networks have an inductive bias towards learning visuallly meaningful features of images faster than noise.
By training a neural network to restore a degraded frame of a video, it can reuse much of what it learned to quickly restore the next degraded frame of the video - all without a dataset!
Login Demo Github. Live website. (2021)
A simple React website with login functionality to get to know how AWS Amplify works.
Statistical Evaluation of CCG Cards (2017)
Many simple Hearthstone cards can be reduced to vectors. We can then use linear regression to measure the mana cost of each "point" of an attribute.
Naïve linear regression turns out to be fairly inaccurate. Building a nonlinear model with prior knowledge of a game then using local search could be one solution. Instead of relying on expert knowledge of the domain, we could also just feed replay data to a neural network. One of these days, I'll return to this topic.