Shuang Liang

News

• 11/26: Talk at the SIAM MDS minisymposium "Mathematics of Generalization in Overparameterized Models" (Scheduled). 

• 08/26: Talk at the Math Machine Learning seminar MPI MIS + UCLA. 

• 07/26: New preprint on the implicit bias and feature learning of SGD in training ReLU networks. 

• 06/26: Talk at the SIAM OP minisymposium "Global Optimization for Neural Networks".  

• 03/26: Talk at the MFO workshop "Modern and Emerging Phenomena in Machine Learning". 

• 01/26: Paper on chaos and fractals in gradient descent optimization accepted to ICLR 2026. 

• 11/25: Awarded the John Fellowship, UCLA. 

• 06/25: Talk at the Scalable Statistical Machine Learning Lab, UCSD. 

• 01/25: Paper on implicit bias of mirror descent in ReLU networks accepted to ICLR 2025. 

Research 

I aim to better understand neural network learning as shaped by the training process. I am excited about: 

• Training dynamics: the evolution of parameters, losses, and related quantities during training; 

• Implicit bias of training algorithms: which predictor the algorithm selects among the many candidates that perform well on the training data; 

• Feature learning: what features the network learns and how they emerge during training; 

• The influence of network architecture, optimizer, parameter initialization, step size, etc. 

Papers 

• Implicit Bias of SGD in Multivariate ReLU Networks: Effective Width Collapse
  Shuang Liang, Tom Jacobs, Guido Montúfar

• Gradient Descent with Large Step Sizes: Chaos and Fractal Convergence Region
  Shuang Liang, Guido Montúfar  
  ICLR 2026

• Implicit Bias of Mirror Flow for Shallow Neural Networks in Univariate Regression
  Shuang Liang, Guido Montúfar
  ICLR 2025 (Spotlight)

• Pull-back Geometry of Persistent Homology Encodings
  Shuang Liang, Renata Turkeš, Jiayi Li, Nina Otter, Guido Montúfar
  TMLR 2024

Miscellaneous