Sharon Li

About

I am an Assistant Professor in the Department of Computer Sciences at the University of Wisconsin Madison. Previously I spent a wonderful year as a postdoc researcher in the Computer Science department at Stanford University, working with Chris Ré. I completed my PhD from Cornell University in 2017, where I was fortunate to be advised by John E. Hopcroft. I've spent time at Google AI as an intern, and Facebook AI as a Research Scientist. I have received Facebook Research Award, JPMorgan early-career faculty award, and was named Forbes 30 Under 30 in Science. I am currently a faculty fellow at the Madison Teaching and Learning Excellence (MTLE) program.

My broad research interests are in deep learning, a branch of machine learning. My time in both academia and industry has shaped my view and approach in research. My research develops algorithms and fundamental understandings to enable reliable open-world learning, which can function safely and adaptively in the presence of evolving and unpredictable data stream. Research topics that I am currently focusing on include:

Out-of-distribution detection for reliable ML;

Open-world machine learning.

News

1/20/2022: Two papers accepted into ICLR 2022.
12/3/2021: Will serve as an area chair for ICML 2022.
12/1/2021: Two papers on OOD detection accepted into AAAI 2022.
11/1/2021: Received Facebook Research Award.
10/25/2021: Received gift funding from Google Brain.
10/22/2021: We released a comprehensive survey on generalized OOD detection.
10/18/2021: Received gift funding from Facebook.
9/28/2021: Three papers on out-of-distribution detection accepted to NeurIPS 2021. Congrats to the team!
7/22/2021: Paper on frequency-domain image translation (FDIT) accepted to ICCV 2021.
7/15/2021: Received Madison Teaching and Learning Excellence Fellowship.
6/18/2021: Paper on outlier mining for out-of-distribution detection is accepted to ECML 2021.
6/4/2021: Received JPMorgan Chase early career outstanding faculty award.
6/2021: Received American Family Funding Intiative Awards.
4/4/2021: Will be co-organizing two ICML'21 workshops on uncertainty quantification.
2/28/2021: Two papers on OOD detection accepted to CVPR 2021 (including one oral presentation).
1/12/2021: Paper on Model Patching (for handling subgroup shift) accepted to ICLR 2021.
10/28/2020: Will serve as a mentor at the Women in Machine Learning (WiML) workshop at NeurIPS 2020.
9/25/2020: Paper on Energy-based Out-of-distribution Detection accepted to NeurIPS 2020.
8/17/2020: Joined the CS Department at UW-Madison as an Assistant Professor. Here is a featured interview article.
05/19/2020: Gave an invited talk on Out-of-distribution Uncertainty Estimation and Robustness in Open-World Machine Learning at Air Force Research Laboratory’s Workshop.
4/23/2020: Selected Young Researcher to participate in the 8th Heidelberg Laureate Forum.
3/14/2020: Will serve as Program Chair for ICML'20 workshop on Uncertainty and Robustness in Deep Learning.
2/26/2020: Wrote a blog series on automating the art of data augmentation, featuring latest works on the practice, theory and new direction of data augmentation.
12/3/2019: My talk is listed as one of 30 Influential AI Presentations in 2019.
12/3/2019: Honored and thrilled to be featured in Forbes 30 Under 30 list in Science.
3/8/2019: Honored to be featured in 30 Under 30 Leading Women in AI.
2/24/2019: Paper on KNN-based robustness accepted as an oral presentation at CVPR 2019.
2/24/2019: Will serve as Program Chair for ICML'19 workshop on Uncertainty and Robustness in Deep Learning.
1/2019: Gave a talk at Deep Learning Summit San Francisco .
5/2/2018: Research featured in a TechCrunch and Wired article.
10/1/2018: Gave an invited talk at Microsoft Research AI in Redmond, WA.
9/26/2018: Gave a talk at Grace Hopper Celebration (GHC) Artificial Intelligence track in Houston, TX.
1/29/2018: Paper on ODIN for detecting out-of-distribution examples in neural networks accepted into ICLR 2018. .

[Services]:

Program chair: ICML workshop on Uncertainty and Robustness in Deep Learning (UDL) 2019 & 2020.

Area chair and senior program committee: NeurIPS, ICLR, ICML and AAAI.

Co-organizer: ICML workshop on Uncertainty and Robustness in Deep Learning (UDL) 2021.

Co-organizer: ICML workshop on Distribution-free Uncertainty Quantification (DFUQ) 2021.

[Recent/upcoming talks]:

3/2022: AI Seminar at Oregon State University

12/2021: NeurIPS'21 workshop on ImageNet: past, present, and future

12/1/2021: PhysicsMeetsML seminar

9/28/2021: Future of Data-Centric AI virtual event

8/27/2021: Talk at Facebook

8/21/2021: Keynote @ Artificial Intelligence for Anomalies and Novelties (AI4AN 2021)

8/22/2021: Keynote @ Weakly-supervised Representation Learning (WSRL 2021 workshop)

8/2/2021: IFDS workshop on Statistical Approaches to Understanding Modern ML Methods

4/27/2021: MINDS & CIS Seminar at John Hopkins University

4/17/2021: Women in Scientific Education and Research (WISER) at UW-Madison

4/1/2021: Open Data Science Conference (ODSC) East 2021

11/4/2020: Women in Computer Science (WiCS) Seminar at Stanford

10/30/2020: Information Systems Seminar at UC Berkeley

10/28/2020: SILO Seminar at UW-Madison. [Recording]

9/21/2020: MLOS Seminar at UW-Madison+Microsoft. [Recording]

[Teaching]:

Fall 2021: CS762 Advanced Deep Learning

Spring 2021: CS540 Introduction to Artificial Intelligence .

Fall 2020: CS839 Advanced Topics in Deep Learning (a newly developed graduate-level course).

Misc
I travel and occasionally take photos. Here is my pictorial Travel Memo. This is the treasure I shoot with.

Research Highlights

VOS: Learning What You Don’t Know by Virtual Outlier Synthesis

Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li

One of the key challenges in OOD detection is that models lack supervision signals from unknown data. Previous approaches rely on real outlier datasets for model regularization, which can be costly and sometimes infeasible to obtain in practice. In this paper, we present VOS, a novel framework for OOD detection by adaptively synthesizing virtual outliers that can meaningfully regularize the model's decision boundary during training. Specifically, VOS samples virtual outliers from the low-likelihood region of the class-conditional distribution estimated in the feature space. VOS achieves strong performance on both object detection and image classification models.

ICLR 2022
PDF
Code
Review

ReAct: Out-of-distribution Detection With Rectified Activations

Yiyou Sun, Chuan Guo and Yixuan Li

This paper provides a simple activation rectification strategy termed ReAct, which truncates thehigh activations during test time for OOD detection. We provide both empirical and theoreticalinsights characterizing and explaining the mechanism by which ReAct improves OOD uncertainty estimation. By rectifying the activations, the outsized contribution of hidden units on OOD outputcan be attenuated, resulting in a stronger separability from ID data. Extensive experiments show ReAct can significantly improve the performance of OOD detection on both common benchmarksand large-scale image classification models.

NeurIPS 2021
PDF
Code
Review

Can multi-label classification networks know what they don’t know?

Haoran Wang, Weitang Liu, Alex Bocchieri and Yixuan Li

Improved methods for OOD detection in multi-class classification have emerged, while OOD detection methods for multi-label classification remain underexplored and use rudimentary techniques. We propose JointEnergy, a simple and effective method, which estimates the OOD indicator scores by aggregating label-wise energy scores from multiple labels. We show that JointEnergy can be mathematically interpreted from a joint likelihood perspective. Our results show consistent improvement over previous methods that are based on the maximum-valued scores, which fail to capture joint information from multiple labels. We demonstrate the effectiveness of our method on three common multi-label classification benchmarks, including MS-COCO, PASCAL-VOC, and NUS-WIDE. We show that JointEnergy can reduce the FPR95 by up to 10.05% compared to the previous best baseline, establishing state-of-the-art performance.

NeurIPS 2021
PDF
Code
Review

On the Importance of Gradients for Detecting Distributional Shifts in the Wild

Rui Huang, Andrew Geng and Yixuan Li

Existing OOD detection approaches primarily rely on the output or feature space for deriving OOD scores, while largely overlooking information from the gradient space. In this paper, we present GradNorm, a simple and effective approach for detecting OOD inputs by utilizing information extracted from the gradient space. GradNorm directly employs the vector norm of gradients, backpropagated from the KL divergence between the softmax output and a uniform probability distribution. Our key idea is that the magnitude of gradients is higher for in-distribution (ID) data than that for OOD data, making it informative for OOD detection. GradNorm demonstrates superior performance, reducing the average FPR95 by up to 10.89% compared to the previous best method.

NeurIPS 2021
PDF
Code
Review

MOS: Towards Scaling Out-of-distribution Detection for Large Semantic Space

Rui Huang and Yixuan Li

Existing OOD detection solutions are mainly driven by small datasets, with low resolution and very few class labels (e.g., CIFAR). As a result, OOD detection for large-scale image classification tasks remains largely unexplored. In this paper, we bridge this critical gap by proposing a group-based OOD detection framework, along with a novel OOD scoring function termed MOS. Our key idea is to decompose the large semantic space into smaller groups with similar concepts, which allows simplifying the decision boundaries between in- vs. out-of-distribution data for effective OOD detection. Our method scales substantially better for high-dimensional class space than previous approaches. We evaluate models trained on ImageNet against four carefully curated OOD datasets, spanning diverse semantics. MOS establishes state-of-the-art performance, reducing the average FPR95 by 14.33% while achieving 6x speedup in inference compared to the previous best method.

CVPR 2021, Oral Presentation
PDF
Code
Review

MOOD: Multi-level Out-of-distribution Detection

Ziqian Lin*, Sreya Dutta Roy* and Yixuan Li

(* indicates equal contributions)

While improved OOD detection methods have emerged, they often rely on the final layer outputs and require a full feedforward pass for any given input. In this paper, we propose a novel framework, multi-level out-of-distribution detection MOOD, which is the first to exploit intermediate classifier outputs for dynamic and efficient OOD inference. We explore and establish a direct relationship between the OOD data complexity and optimal exit level, and show that easy OOD examples can be effectively detected early without propagating to deeper layers. We extensively evaluate MOOD across 10 OOD datasets spanning a wide range of complexities. Experiments demonstrate that MOOD achieves up to 71.05% computational reduction in inference, while maintaining competitive OOD detection performance.

CVPR 2021
PDF
Code
Review

Energy-based Out-of-distribution Detection

Weitang Liu, Xiaoyun Wang, John Owens and Yixuan Li

Determining whether inputs are out-of-distribution (OOD) is an essential building block for safely deploying machine learning models in the open world. However, previous methods relying on the softmax confidence score suffer from overconfident posterior distributions for OOD data. We propose a unified framework for OOD detection that uses an energy score, which can better distinguish in- and out-of-distribution samples than the traditional approach using the softmax scores. Within this framework, energy can be flexibly used as a scoring function for any pre-trained neural classifier as well as a trainable cost function to shape the energy surface explicitly for OOD detection. On a CIFAR-10 pre-trained WideResNet, using the energy score reduces the average FPR (at TPR 95%) by 18.03% compared to the softmax confidence score.

NeurIPS 2020
PDF
Code
Review

Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks

Shiyu Liang, Yixuan Li and 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 34.7% to 4.3% on the DenseNet (applied to CIFAR-10) when the true positive rate is 95%.

ICLR 2018
PDF
Code

Sharon Yixuan Li

About

News

Research Highlights

Sponsors