I am Yixuan Li (pronounced as "e-shwen-lee"). I am a Research Scientist at Facebook Research. Before joining Facebook, I finished my Ph.D. in ECE from Cornell University, advised by John E. Hopcroft. My thesis committee members are Kilian Q. Weinberger and Thorsten Joachims. A key focus of my research has been on computer vision and deep learning. Recent projects include deep learning interpretability, optimizing neural networks with efficient computational cost, adversarial training of deep generative models, improving neural network reliability, and theoretical aspect of deep learning.

Prior to Cornell, I graduated from Shanghai Jiaotong University with B.Eng in Information Engineering in 2013. I spent two summers at Google Research Mountain View in 2015 and 2016. I spend summer in 2017 as a machine learning scientist intern at GrokStyle, building cutting-edge visual search technologies with deep learning and computer vision.

I travel and occasionally take photos. Here is my pictorial Travel Memo.

Update (5/16/2018): Selected to speak at Grace Hopper Celebration (GHC) Artificial Intelligence track in September.
Update (5/12/2018): Paper on understanding the loss surface of neural networks accepted into ICML 2018.
Update (4/3/2018): Received CVPR'18 Doctoral Consortium travel award.
Update (3/13/2018): Served on a panel at Facebook's Women in Research Lean In (WiRL) Circle.
Update (1/29/2018): Paper on Neural Network Reliability accepted into ICLR 2018.
Update (11/25/2017): Received ACM-W Scholarship in 2017.
Update (10/16/2017): I will be presenting at Women in Machine Learning (WiML) workshop in December this year.
Update (10/2/2017): Successfully defended my thesis. Slides available here.
Update (9/6/2017): I will be joining Facebook as a full-time Research Scientist in October 2017.
Update (8/5/2017): Selected as one of the Rising Stars in EECS 2017.
Update (6/6/2017): Paper accepted for publication in Transactions on Knowledge Discovery from Data (TKDD).
Update (5/16/2017): I will be speaking at Grace Hopper Celebration (GHC) Artificial Intelligence track in October 2017.
Update (3/12/2017): Received ICLR 2017 Student Travel Award.
Update (2/27/2016): Paper on StackedGAN has been accepted into CVPR 2017.
Update (2/6/2017): Paper on Snapshot Ensembles has been accepted into ICLR 2017.
Update (12/20/2016): My summer internship paper at Google Research is invited to the industrial track in WWW 2017.
Update (2/5/2016): I will be interning at Machine Intelligence at Google Research (Mountain View) for the summer. I am very excited about it!
Update (2/4/2016): Paper on Convergent Learning has been accpeted for oral presentation (5.7%) in ICLR 2016! (check out preprint here)

[CV] [Google Scholar]

Research Highlights

Exploring the Limits of Weakly Supervised Pretraining.
D. Mahajan, R.B. Girshick, V. Ramanathan, K. He, M. Paluri, Y. Li, A. Bharambe, and L.J.P. van der Maaten.
ImageNet classification is the de facto pretraining task for these models. Yet, ImageNet is now nearly ten years old and is by modern standards "small". In this paper, we present a unique study of transfer learning with large convolutional networks trained to predict hashtags on billions of social media images. Our experiments demonstrate that training for large-scale hashtag prediction leads to excellent results. We show improvements on several image classification and object detection tasks, and report the highest ImageNet-1k single-crop, top-1 accuracy to date: 85.4% (97.6% top-5). We also perform extensive experiments that provide novel empirical data on the relationship between large-scale pretraining and transfer learning performance.
Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks.
Shiyu Liang, Yixuan Li, R. Srikant.
Modern neural networks are known to generalize well when the training and testing data are sampled from the same distribution. However, when deploying neural networks in real-world applications, there is often very little control over the testing data distribution. For example, a neural network trained for classifying handwritten digits can assign high confidence labels to images of animal. This has raised great concerns in AI Safety, in particular how can classifiers obtain awareness of uncertainty when shown new kinds of inputs, i.e., out-of-distribution examples. We propose ODIN for detecting out-of-distribution images, which can be applied on any pre-trained neural networks. ODIN reduces the false positive rate from the baseline approach (Hendrycks & Gimpel, 2017) 34.7% to 4.3% on the DenseNet (applied to CIFAR-10) when the true positive rate is 95%.
Snapshot Ensembles: Train 1 Get M for Free.
Gao Huang*, Yixuan Li*, Geoff Pleiss, Zhuang Liu, John Hopcroft, Kilian Weinberger. (* equal contribution)
Training multiple deep networks for model averaging is computationally expensive. In this paper, we propose a method to obtain the seemingly contradictory goal to obtain ensembles of multiple neural network at no additional training cost. We achieve this goal by letting a single neural network converge into several local minima along its optimization path and save the model parameters. To obtain repeated rapid convergence we leverage recent work on cyclic learning rate schedules. The resulting technique, which we refer to as Snapshot Ensembling, is surprisingly simple, yet effective. On Cifar-10 and Cifar-100 our DenseNet Snapshot Ensembles obtain error rates of 3.4% and 17.4% respectively.
Measuring and Inferring User Interest from Gaze.
Yixuan Li, Pingmei Xu, Dmitry Lagun and Vidhya Navalpakkam.
In this paper, we investigate the relationship between mobile users' implicit interest inferred from attention metrics, such as eye gaze or viewport time, and explicit interest expressed by users. We present the first quantitative gaze tracking study using front-facing camera of mobile devices instead of specialized, expensive eye-tracking devices. In the context of Google Play Store pages, we find significantly different distribution of gaze metrics on items that users rate as interesting vs. not. We leverage this insight by building a prediction model that is able to infer a user's interest ratings from the the non-click actions of the user. Our model is able to attain AUC of 90.32% in predicting user interest at an individual item level.
Stacked Generative Adversarial Networks
Xun Huang, Yixuan Li, Omid Poursaeed, John Hopcroft, Serge Belongie.
In this paper we propose a novel generative model named Stacked Generative Adversarial Networks (SGAN), which is trained to invert the hierarchical representations of a discriminative bottom-up deep network. Our model consists of a top-down stack of GANs, each trained to generate "plausible" lower-level representations, conditioned on higher-level representations. A representation discriminator is introduced at each feature hierarchy to encourage the representation manifold of the generator to align with that of the bottom-up discriminative network, providing intermediate supervision. Unlike the original GAN that uses a single noise vector to represent all the variations, our SGAN decomposes variations into multiple levels and gradually resolves uncertainties in the top-down generative process. Experiments demonstrate that SGAN is able to generate diverse and high-quality images, as well as being more interpretable than a vanilla GAN.
Convergent Learning: Do different neural networks learn the same representations?
Yixuan Li*, Jason Yosinski*, Jeff Clune, Hod Lipson and John Hopcroft. (* equal contribution)
In this paper we investigate the extent to which neural networks exhibit what we call convergent learning, which is when the representations learned by multiple nets converge to a set of features which are either individually similar between networks or where subsets of features span similar low-dimensional spaces. We propose a specific method of probing representations: training multiple networks and then comparing and contrasting their individual, learned representations at the level of neurons or groups of neurons. This initial investigation reveals a few previously unknown properties of neural networks. The insights described here include (1) that some features are learned reliably in multiple networks, yet other features are not consistently learned; (2) units learn to span low-dimensional subspaces and, while these subspaces are common to multiple networks, the specific basis vectors learned are not.
In a World That Counts: Clustering and Detecting Fake Social Engagement at Scale
Yixuan Li, Oscar Martinez, Xing Chen, Yi Li, John Hopcroft.
How can web services that depend on user generated content discern fake social engagement activities by spammers from legitimate ones? In this paper, we focus on the social site of YouTube and the problem of identifying bad actors posting inorganic contents and inflating the count of social engagement metrics. We propose an effective method, Leas (Local Expansion at Scale), and show how the fake engagement activities on YouTube can be tracked over time by analyzing the temporal graph based on the engagement behavior pattern between users and YouTube videos. We demonstrate the effectiveness of our deployment at Google by achieving an manual review accuracy of 98% on YouTube Comments graph in practice. Leas is actively in use at Google, searching for daily deceptive practices on YouTube's engagement graph spanning over a billion users.
Uncovering the Small Community Structure in Large Networks: A Local Spectral Approach.
Yixuan Li, Kun He, David Bindel, John Hopcroft.
Large graphs arise in a number of contexts and understanding their structure and extracting information from them is an important research area. Early algorithms on mining communities have focused on the global structure, and often run in time functional to the size of the entire graph. In this paper, we propose a novel approach for finding overlapping communities called LEMON (Local Expansion via Minimum One Norm). Different from PageRank-like diffusion methods, LEMON finds the community by seeking a sparse vector in the span of the local spectra such that the seeds are in its support. We show that LEMON can achieve the highest detection accuracy among state-of-the-art proposals. The running time depends on the size of the community rather than that of the entire graph.

Publications

  • Understanding the Loss Surface of Neural Networks for Binary Classification.
    Shiyu Liang, Ruoyu Sun, Yixuan Li, and R. Srikant
    Proceedings of International Conference on Machine Learning (ICML 2018).
    Stockholm, Sweden, July 2018. [PDF]

  • Exploring the Limits of Weakly Supervised Pretraining.
    Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens van der Maaten
    Preprint on arXiv, cs.CV 1805.00932, 2018.
    [PDF][Blog Post] [Github]

  • Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks.
    Shiyu Liang, Yixuan Li, and R. Srikant
    Proceedings of International Conference on Learning Representation (ICLR 2018).
    Vancouver, Canada, April 2018. [PDF][code]

  • Snapshot Ensembles: Train 1, Get M for Free.
    Gao Huang*, Yixuan Li*, Geoff Pleiss, Zhuang Liu, John Hopcroft and Kilian Weinberger
    Proceedings of International Conference on Learning Representation (ICLR 2017).
    Toulon, France, April 24 - 26, 2017.[PDF][code]

  • Stacked Generative Adversarial Networks.
    Xun Huang, Yixuan Li, Omid Poursaeed, John Hopcroft and Serge Belongie
    Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR 2017).
    Honolulu, Hawaii, July 22 - 25, 2017.[PDF][code]

  • Towards Measuring and Inferring User Interest From Gaze.
    Yixuan Li, Pingmei Xu, Dmitry Lagun and Vidhya Navalpakkam
    Accepted to the 26th International Conference on World Wide Web (WWW 2017). [PDF][slides]

  • Convergent Learning: Do different neural networks learn the same representations?
    Yixuan Li*, Jason Yosinski*, Jeff Clune, Hod Lipson and John Hopcroft
    Proceedings of International Conference on Learning Representation (ICLR 2016).
    San Juan, Puerto Rico, May 2016 (Oral presentation 5.7%) [PDF][code][video]

  • In a World that Counts: Clustering and Detecting Fake Social Engagement at Scale.
    Yixuan Li, Oscar Martinez, Xing Chen, Yi Li and John Hopcroft
    Proceedings of the 25th International World Wide Web Conference (WWW 2016).
    Montreal, Canada, April 2016. [PDF][news coverage]

  • The Lifecycle and Cascade of WeChat Social Messaging Groups.
    Jiezhong Qiu, Yixuan Li, Jie Tang, Zheng Lu, Hao Ye, Bo Chen, Qiang Yang and John Hopcroft
    Proceedings of the 25th International World Wide Web Conference (WWW 2016)
    Montreal, Canada, April 2016. [PDF]

  • Deep Manifold Traversal: Changing Labels with Convolutional Features.
    Jacob Gardner*, Paul Upchurch*, Matt Kusner, Yixuan Li, Kilian Weinberger and John Hopcroft
    Preprint on arXiv. November, 2015. [PDF]

  • Scalable and Robust Local Community Detection via Adaptive Subgraph Extraction and Diffusions.
    Kyle Kloster and Yixuan Li
    Preprint on arXiv, cs.SI:1611.05152, 2016. [PDF]

  • Local Spectral Clustering for Overlapping Community Detection.
    Yixuan Li, Kun He, Kyle Kloster, David Bindel and John Hopcroft
    In ACM Transactions on Knowledge Discovery from Data (TKDD), June 2017, to appear. [PDF]

  • Detecting Overlapping Communities from Local Spectral Subspaces.
    Kun He, Yiwei Sun, David Bindel, John Hopcroft, Yixuan Li
    IEEE International Conference on Data Mining (ICDM 2015)
    Atlantic City, NJ, USA. November, 2015 (acceptance ratio: 18.2%) [pdf][full version]

  • Uncovering the Small Community Structure in Large Networks: A Local Spectral Approach.
    Yixuan Li, Kun He, David Bindel, John Hopcroft
    Proceedings of the 24th International World Wide Web Conference (WWW 2015)
    Florence, Italy. May, 2015 (acceptance ratio: 14.1%) [PDF][code]

  • Multicast Capacity with Max-Min Fairness for Heterogeneous Networks.
    Yixuan Li, Qiuyu Peng, Xinbing Wang
    In IEEE/ACM Transactions on Networking (TON), 2014. [PDF]

  • On Multicast Capacity and Delay in Cognitive Radio Mobile Ad-hoc Networks.
    Jinbei Zhang, Yixuan Li, Zhuotao Liu, Fan Wu, Feng Yang, Xinbing Wang
    IEEE Transactions on Wireless Communications, 2015. [PDF]

  • Awards

  • Cornell Graduate Student Conference Grant (WWW’15, ICDM’15), Cornell University, 2015.

  • Cornell University Graduate Fellowship, Graduate School of Cornell University, 2013.

  • Academic Excellence Scholarship (5%), Shanghai Jiao Tong University, 2011 - 2013.

  • National Scholarship of China (3%), Ministry of Education of the People's Republic of China, 2012 - 2013.

  • Meritorious Winner in the American Interdisciplinary Contest in Modeling, COMAP, the Consortium for Mathematics and Its Applications, 2012.

  • First Prize in Chinese Undergraduate Mathematical Contest in Modeling, China Society for Industrial and Applied Mathematics (CSIAM), 2011.

  • Undergraduate Scholarship for Studying Abroad, Shanghai Jiaotong University, 2011.

  • Wen-Yuan Pan Scholarship (1 out of 105), 2010.

  • Work Experiences

  • Machine Learning Scientist Intern, GrokStyle Inc. Ithaca, NY, 2017.5 - 2017.9

  • PhD Research Intern, Google Inc. Mountain View, CA, 2016.5 - 2016.8

  • PhD Research Intern, Google Inc. Mountain View, CA, 2015.5 - 2015.8

  • Talks

  • Towards Understanding the Inner Workings of Deep Neural Networks.
    Grace Hopper Celebration 2017, Artificial Intelligence track, 2017.10.4 [slides]

  • Deep Neural Networks for Visual Recognition: Efficiency, Transparency and Reliability.
    PhD defense talk, 2017.10.2 [slides]

  • Towards Understanding, Improving and Scaling Learning in Deep Neural Networks.
    Invited Talk at Computer Vision Group @ Cornell Tech, 2017.1.25 [slides]

  • Scale, Improve and Understand Learning Through Subspace Embedding.
    A-Exam (thesis proposal exam), 2016.8.22 [slides]

  • Convergent Learning: Do different neural networks learn the same representations?
    Guest Lecture CS4850 Mathematical Foundations for the Information Age, 2016.5.6 [slides]

  • In a World That Counts: Clustering and Detecting Fake Social Engagement at Scale.
    Internaltional World Wide Web Conference (WWW 2016), Montreal, Canada, 2016.4.13 [slides]

  • Convergent Learning: Do different neural networks learn the same representations?
    Invited Talk at Cornell Statistics Student Seminar, 2016.3.22 [slides]

  • Local Spectral Graph Clustering at Scale: Principle and Its Application.
    Invited Talk at Google Research NYC, 2016.2.9 [slides]

  • Convergent Learning: Do different neural networks learn the same representations?
    NIPS'15 Workshop on Feature Extraction, Montreal, Canada, 2015.12.11 [slides][video]

  • Convergent Learning: Do different neural networks learn the same representations?
    Cornell Machine Learning Discussion Group (MLDG), Ithaca, NY, 2015.12.2 [slides]

  • Uncovering the Small Community Structure in Large Networks: A Local Spectral Approach.
    Internaltional World Wide Web Conference (WWW 2015), Florence, Italy, 2015.5.20 [slides]

  • Teaching

  • SP15 CS4850 Mathematical Foundations for the Information Age.

  • Misc

    Here are the two student organizations at Cornell I actively involved in:

  • Vice President, Chinese Students and Scholars Association at Cornell (CSSA). SP15 - SP16

  • Vice President, Technology Entrepreneurship at Cornell (TEC) . FA14 - FA16

  • Cornell at a Glance

    Photos by Yixuan

    Contact