Wenzhi graduated from the University of Liverpool with a BSc in Mathematics. He later completed his master’s in fluid dynamics at Imperial College, where his dissertation was titled ‘Auxiliary Functions for Stochastic System with Weak Noise.’ Afterward, he worked as a research assistant at Hong Kong Baptist University, focusing on optimization. His interests lie in numerical methods, optimization, and fluid dynamics, he is also keen on applying math to real-world problems.

Outside of mathematics, he enjoys jogging (5 km is preferable), travelling (2 landmarks at most per day), and cooking (not good at it, though).

Project title:
Scalable, efficient and compression-aware optimization for deep learning

Supervisor(s):
Michael Murray, Sergey Dolgov

Project description:
Deep neural networks often contain far more parameters than training samples, so minimizing training loss alone should lead to overfitting. Yet in practice, even models trained with simple stochastic gradient methods can interpolate the data while still generalizing well. This paradox is often attributed to implicit regularization, where training dynamics bias models toward low-complexity and compressible solutions. Building on this idea, recent algorithms aim not only to minimize empirical loss but also to prefer flat or compressible minima, achieving improved generalization in some cases even though their mechanisms remain poorly understood. This project combines mathematical analysis with targeted experiments to study these algorithms, characterize their interactions with architectures, training methods, and hyperparameters, and develop compression-aware optimizers that better exploit the generalization potential of neural networks in large-scale distributed settings.