Weihang Zheng

Formulated DAG extraction from Convolution Neural Networks with up to hundreds of millions of weighted parameters. Improved Knowledge Distillation weight transfer during Neural Architecture Search to speed up performance by 50%. Spearheaded and presented the Uniform Crossover Genetic Algorithm for Channel Size Optimization idea to the entire research group, reducing ResNet34’s total parameter size by 33% while achieving higher test accuracy on CIFAR100. Co-authored one publication, CONetV2, under Professor Kostas Plataniotis, accepted for oral presentation at ICMLA.

2D Atari game where a stationary spaceship shoots at moving asteroids coming at it. Given game implementation with C++ linked lists. Programmed spaceship with hitbox detection, trigometry angle approximation, and danger priorization to shoot at asteroids. Beat 2019’s 1st place score of 35k by 5-6x with a score of 190k .

Developed in C an AB-pruning Minimax bot with Zobrist hashing and transpositions tables that explores 10000+ states. Achieved 6th place in a class of 420+ students, 75 of whom beat the professor’s programs to make the final leaderboard.

Generated large-scale Google-maps-inspired UI/backend program that displays 10000+ intersections in any major city. Placed 7th in class for a Travelling Salesman-related NP-C challenge built on top of the map using techniques such as A*/Dijkstra Search, Simulated Annealing, Multithreading, 2-opt; beat the TA program by 10%. Collaborated with two other students for this project selected to present to the following year ECE students.

Created animated impact detection game in C that takes as input user keystrokes. The game is implemented by tracking different states as a finite state machine.