Hongyang R. Zhang Assistant professor of Computer Science Northeastern University hrzhang@northeastern.edu

Hi! I'm Hongyang (Ryan) Zhang. My research interests lie in the intersection of machine learning, algorithms and statistics. Current focus includes neural networks and its theory, weak supervision such as data augmentation, multi-task learning, and network analysis.

I received my Ph.D. in Computer Science from Stanford University, advised by Ashish Goel and Greg Valiant. I did my undergraduate study in Computer Science from the ACM Honored Class at Shanghai Jiao Tong University. Here is my resume and Google Scholar profile.

Recent Updates

• Sep 2020: I'm teaching an advanced course in deep learning this semester! Feedback more than welcome.

• July 2020: At COLT’20, we show an interesting dynamic of gradient descent for learning over-parametrized two-layer ReLU neural nets. We found that gradient descent first learns lower-order tensors/moments and then learns higher-order tensors/moments (Yuanzhi's video)!

• May 2020: At ICML’20, we describe a formal analysis of data augmentation. Inspired by the analysis, we propose an uncertainty-based sampling scheme that achieves SoTA results (Sen's code release)!

• Feb 2020: Blog posts that illustrate our recent progress on multi-task learning and data augmentation! The data augmentation blog post is part of Sharon Y. Li's expository series about exciting recent progress in this direction.

• Dec 2019: Our work on multi-task learning is accepted at ICLR’20. We rigorously formulate the notion of information transfer in simple models. This leads to several insights with practical implications, which are validated on the SoTA BERT model.

Publications

(asterisk indicates alphabetical authorship or equal contribution)

• Learning Over-Parametrized Two-Layer ReLU Neural Networks beyond NTK
  with Yuanzhi Li and Tengyu Ma
  Conference on Learning Theory (COLT) 2020
  

• On the Generalization Effects of Linear Transformations in Data Augmentation
  Sen Wu*, H. R. Zhang*, Gregory Valiant and Christopher Ré
  International Conference on Machine Learning (ICML) 2020
  [blog]

• Understanding and Improving Information Transfer in Multi-Task Learning
  Sen Wu*, H. R. Zhang* and Christopher Ré
  International Conference on Learning Representations (ICLR) 2020
  [blog] [video] [arxiv]

• Pruning based Distance Sketches with Provable Guarantees on Random Graphs
  H. R. Zhang, Huacheng Yu and Ashish Goel
  The Web Conference (WWW) 2019 (oral presentation)

• Recovery Guarantees for Quadratic Tensors with Limited Observations
  H. R. Zhang, Vatsal Sharan, Moses Charikar and Yingyu Liang
  International Conference on AI and Statistics (AISTATS) 2019

• Algorithmic Regularization in Over-parameterized Matrix Sensing and Neural Networks with Quadratic Activations
  with Yuanzhi Li and Tengyu Ma
  Conference on Learning Theory (COLT) 2018 (Best Paper Award)

• Approximate Personalized PageRank on Dynamic Graphs
  H. R. Zhang, Peter Lofgren and Ashish Goel
  KDD 2016 (oral presentation) [code]

• Incentives for Strategic Behavior in Fisher Market Games
  with Ning Chen, Xiaotie Deng, Bo Tang
  AAAI 2016

• A Note on Modeling Retweet Cascades on Twitter
  with Ashish Goel, Kamesh Munagala, Aneesh Sharma
  Workshop on Algorithms and Models for the Web Graph (WAW) 2015

• Connectivity in Random Forests and Credit Networks
  with Ashish Goel, Sanjeev Khanna, Sharath Raghvendra
  Symposium on Discrete Algorithms (SODA) 2015

• Computing the Nucleolus of Matching, Cover and Clique Games
  with Ning Chen and Pinyan Lu
  AAAI 2012 (oral presentation)

• Incentive Ratios of Fisher Markets
  with Ning Chen, Xiaotie Deng, Jie Zhang
  International Colloquium on Automata, Languages, and Programming (ICALP) 2012

• On Strategy-proof Allocation without Payments or Priors
  with Li Han, Chunzhi Su, Linpeng Tang
  Conference on Web and Internet Economics (WINE) 2011

Blog posts

• When Multi-Task Learning Works – And When It Doesn’t
  based on joint work with Sen Wu and Chris Ré.

• Automating the Art of Data Augmentation – Part III Theory
  edited with Sharon Y. Li and Chris Ré.

Teaching

CS 7140, Advanced Machine Learning, Spring 2021

CS 7180, Special Topics in AI, Fall 2020: Algorithmic and Statistical Aspects of Deep Learning.

Misc

My PhD thesis has studied two broad sets of questions:

• Generalization in over-parameterized models: Non-convex methods are the method of choice for training most ML models in practice. How much data is needed to train a model that generalizes well to unseen data?

• Algorithms for large-scale social networks: Dealing with large-scale social network data requires methods that scale to tens of millions of users. In the meantime, real world social networks admit structural properties. How do we exploit their structures to provide better algorithms and guarantees?

My Chinese name is 张泓洋 written in Chinese characters.