Tian Li
@litian0331
Followers
1K
Following
713
Media
19
Statuses
158
Assistant Professor @UChicagoCS @DSI_UChicago | PhD @CSDatCMU
Chicago, IL
Joined March 2017
I am taking new Ph.D. students from @UChicagoCS and @DSI_UChicago in the 2024-2025 cycle! If you are interested in distributed optimization, data sharing, and trustworthy ML, please feel free to apply! More info on our research:
2
91
324
Personal update: I am excited to join the University of Chicago @UChicago @UChicagoCS @DSI_UChicago as an Assistant Professor starting in Summer 2024! Looking forward to what’s to come!.
44
13
358
[#ICLR2021 Camera-Ready].We propose Tilted Empirical Risk Minimization (TERM), a general and flexible extension to ERM framework. TERM increases or decreases the influence of outliers, by tuning a tilt hyperparameter ‘t’ to enable fairness or robustness:.
1
21
100
Happy to be named as Rising Stars in ML by @ml_umd !.
Check out and follow amazing works by our 2021 "Rising Stars in ML":.Xinyun Chen @xinyun_chen_ (UC Berkeley): Qi Lei @Qi_Lei_ (Princeton): Tian Li @litian0331 (CMU): More info:
4
3
39
#ICML2021. Ditto: Fair and Robust Federated Learning Through Personalization. Joint work with Shengyuan, @abeirami, and @gingsmith. Ditto also won the best paper award at Secure ML Workshop at ICLR (.
1
5
35
I am also actively looking for students (PhD/masters/undergrads) and postdocs in the areas of optimization, trustworthy learning, and distributed/federated learning. Feel free to reach out if you’re interested.
7
3
20
Describing our recent tilted empirical risk minimization work in a <10-min read👇.
Empirical Risk Minimization is a popular approach, but it may be unfair and susceptible to outliers/class imbalance. The new post by Tian Li proposes a fix! . Post: .Paper: Code:
0
1
15
I will be at NeurIPS the whole week. Just email me if you’d like to chat and learn more. 😀.
I am taking new Ph.D. students from @UChicagoCS and @DSI_UChicago in the 2024-2025 cycle! If you are interested in distributed optimization, data sharing, and trustworthy ML, please feel free to apply! More info on our research:
0
0
12
Come to our poster session on May 3rd (Mon) 5-7 pm PDT to chat about TERM. Also, the video is available at:.
[#ICLR2021 Camera-Ready].We propose Tilted Empirical Risk Minimization (TERM), a general and flexible extension to ERM framework. TERM increases or decreases the influence of outliers, by tuning a tilt hyperparameter ‘t’ to enable fairness or robustness:.
0
1
11
Before that, I will spend one year as a postdoctoral researcher at Meta FAIR.
2
0
7
Thanks for hosting @ml_umd ! Had great fun virtually visiting there.
PhD student Tian Li @CSDatCMU concluded our Rising Stars in Machine Learning speaker series with her talk "On Heterogeneity in Federated Settings." . Thanks for sharing your engaging research with us @litian0331!
0
0
7
please considering submitting! deadline is April 12.
Inviting you to submit to our Federated Learning Systems (FLSys) Workshop @ MLSys 2023: See you in Miami.
1
1
7
I am deeply grateful for everyone who has helped and supported me along this journey, especially my advisor @gingsmith and my parents!.
1
0
5
Thank you for hosting us. Really enjoyed #RisingStarsinEECS2022!.
Thank you to all who participated in #RisingStarsinEECS2022! @utexasece and @UTCompSci were proud to host these amazing future leaders in academia. #WhatStartsHere
0
0
5
TERM recovers a family of objectives parameterized by t, and allows a smooth transition from min-loss, to avg-loss, and max-loss. We also show TERM approximates a popular family of quantile losses (e.g., median loss), which can be useful in practice but hard to directly optimize.
1
0
4
We theoretically analyze fairness and robustness of Ditto for a class of linear problems, and empirically demonstrate that Ditto outperforms strong robust or fair baselines across a set of benchmarks.
1
0
2
TERM outperforms SOTA robust regression baselines (both for feature noise and label noise), particularly in noise regimes
1
0
2
Congrats!! BU is lucky to have you!.
Some personal news in this mess: I'm delighted to join @BU_ece in @BUCollegeofENG as an assistant professor starting Jan 2021. I'm very grateful to all my mentors, family, and friends who kindly supported me through this turbulent job search.
0
0
2
We show that TERM can be optimized by simple tweaks to ERM optimization framework. We develop batch and stochastic first-order optimization methods for solving TERM. The training times for TERM ran within 2x of ERM in all our experiments. The code:
1
0
2
1
0
2
We use TERM to address a multitude of ML applications, showing that it can achieve superior/competitive performance with state-of-the-art, problem-specific solutions. E.g., robust regression, robust classification, fair PCA, handling class imbalance, mitigating noisy annotators.
1
0
2
For fair PCA, TERM can allow for a flexible tradeoff between performance and fairness, and recovers min-max solutions with a large t
1
0
2
To deploy FL in the real world, beyond optimizing for accuracy, the resulting systems must also satisfy a number of constraints such as fairness, robustness, and privacy.
1
0
2
TERM is competitive with state-of-the-art methods for classification with imbalanced classes
1
0
2
the deadline has been extended. please consider submitting!.
We updated our CFP and important dates! .Ddl for regular workshop papers: Mar 7, 2022.Ddl for submissions w/ ARR reviews: Mar 21, 2022 .We also accept relevant published papers that will benefit from exposure to the audience of the workshop. More info:
0
0
1
@AlexKale17 @UChicago @UChicagoCS @DSI_UChicago Thank you, Alex! Same here—also looking forward to working with you.
0
0
1
@abeirami @FarisSbahi @quantumVerd @meisamrr yes, that converges similarly as SGD. another interesting question could be to understand the generalization of TERM in non-convex settings, i.e., whether a small positive t can lead to flat local minima, or what’s a good way to think about it (given its relations with robust opt.
0
0
1
See our paper and code below for complete theoretical and experimental results. Preprint: .Code:
0
0
1
0
0
1
@KWhiteheadLab @vyas_sekar @CarnegieMellon interesting! looking forward to my hooding (F-F) (several years later :) ).
0
0
1
@DimitrisPapail @UChicago @UChicagoCS @DSI_UChicago Thanks for your kind words! +1 Definitely hope to collaborate more (and I will be close by). 😃.
0
0
1
We first show that fairness (representation disparity) and robustness (against poisoning attacks) constraints can be directly competing with each other---Fair methods are not robust, and robust methods are not fair.
1
0
1
0
0
1
Complete theoretical and empirical results are presented in the paper. Joint work with @abeirami, Maziar, and @gingsmith. See you at ICLR online!.
0
0
1
To solve Ditto, we alternate between solving for a global model and solving for personalized models. Ditto can be viewed as a lightweight personalized add-on on top of any global solver.
1
0
1
To simultaneously address these constraints, we propose Ditto as a lightweight global-regularized MTL objective to produce personalized models for FL. Despite its simplicity, the benefits of Ditto go beyond accuracy---it is inherently more robust and more fair than SOTA baselines
1
0
1
0
0
1
@kenziyuliu @CarnegieMellon @Stanford @gingsmith @zstevenwu @SCSatCMU @KairouzPeter @GoogleAI congrats!.
1
0
1