Hongyi Zhang (张宏毅)

I am a Research Scientist in the Applied Machine Learning (AML) team at ByteDance. We are hiring!

I recently graduated from MIT. During my PhD I had a great time working with Prof. Suvrit Sra on a bunch of interesting Riemannian optimization problems. I also joyfully spent two summers at FAIR, mixup and fixup neural networks.

I was also affiliated with LIDS and the Machine Learning Group at MIT.

Innovations today shape our future. My personal take on the most important challenges are: 1) understanding, engineering and enhancing intelligence; 2) high quality pre-college education that is scalable and sustainable.

You can find me at firstname (dot) lastname (dot) pku (at) gmail (dot) com.

Education

2013 - 2019 Massachusetts Institute of Technology
Ph.D. in Cognitive Science
2008 - 2013 Peking University
Bachelor in Machine Intelligence

Experience

2017/2018 Summer Research Intern,
at Facebook AI Research with Yann Dauphin
2015/2016 Fall Teaching Assistant,
MIT 9.520/6.860 -- Statistical Learning Theory and Applications
2014 Fall Teaching Assistant,
MIT 9.66/6.804/9.660 -- Computational Cognitive Science
2012 Summer Research Intern,
at TTI Chicago with Prof. Raquel Urtasun

Publications

Fixup Initialization: Residual Learning Without Normalization. [OpenReview]
Hongyi Zhang*, Yann N. Dauphin*, Tengyu Ma.
7th International Conference on Learning Representations (ICLR 2019).

An Estimate Sequence for Geodesically Convex Optimization. [arXiv]
Hongyi Zhang, Suvrit Sra.
Proceedings of the 31st Conference On Learning Theory, PMLR 75:1703-1723, 2018.

mixup: Beyond Empirical Risk Minimization. [paper][code]
Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, David Lopez-Paz.
6th International Conference on Learning Representations (ICLR 2018)

Matrix Completion from O(n) Samples in Linear Time. [Paper][Long version]
David Gamarnik, Quan Li, Hongyi Zhang;
Proceedings of the 30th Conference on Learning Theory, PMLR 65:940-947, 2017.

Riemannian SVRG: Fast Stochastic Optimization on Riemannian Manifolds. [Paper]
Hongyi Zhang, Sashank J. Reddi, Suvrit Sra.
30th Conference on Neural Information Processing Systems (NIPS 2016)

First-order Methods for Geodesically Convex Optimization. [Paper]
Hongyi Zhang, Suvrit Sra.
Proceedings of the 29th Conference on Learning Theory, PMLR 49:1617-1638, 2016.

Physics 101: Learning Physical Object Properties from Unlabeled Videos. [Paper][Project]
Jiajun Wu, Joseph Lim, Hongyi Zhang, Joshua Tenenbaum and William Freeman.
Proceedings of the British Machine Vision Conference 2016 (BMVC 2016)

Writing Customized Proposals for Probabilistic Programs as Probabilistic Programs.
Hongyi Zhang, Vikash Mansinghka.
3rd NIPS Workshop on Probabilistic Programming. 2014.

Understanding High-Level Semantics by Modeling Traffic Patterns. [Paper][Supplemental][Video]
Hongyi Zhang, Andreas Geiger, Raquel Urtasun.
International Conference on Computer Vision (ICCV'13)