I guess this announcement is a bit (or super) late. I am on the 2023-2024 faculty job market for CS and data science!! Here is my humble homepage: zhundeng.org/home

First off, I should mention this was joint work with a lot of excellent folks who know a lot more about constrained RL than I: Konwoo Kim, Zuxin Liu, Sanjiban Choudhury, Steven Wu, Ding Zhao. Second, if you prefer video, check out our explanation here: youtube.com/watch?v=2jFMBB… [2/n]

PCR is a popular technique for error-in-variables regression, a generalization of linear regression in which the observed covariates are corrupted with noise. At #NeurIPS23 we provide the first time-uniform finite sample guarantees for PCR whenever data is collected adaptively.

Algorithmic decision-making often involves assigning decisions to agents with incentives to strategically modify their input to the algorithm. Furthermore, in many domains of interest the decision-maker only observes feedback for rounds in which they assign a positive decision.

I'm excited to be on the PC for FORC24 (Foundations of Responsible Computing), one of my favorite new venues in CS! It's a great venue for submitting rigorous work on topics like privacy, fairness, robustness, etc. CfP is out, so submit your best work! responsiblecomputing.org/forc-2024-call…

Come to the robot learning workshop (Hall B2) at #NeurIPS2023 to hear what I think the right algorithm for imitation learning for robots is & how it took me 1+ years to write a 1pg. proof. Joint work w/ an incredible ugrad Juntao Ren + Steven Wu, Sanjiban Choudhury, and Drew Bagnell.

Excited for the next talk in our #NetworkInequality lecture series! Join Luke Guerdan on Feb 29 at 2pm CET for "Human-Algorithm Decision-Making Under Imperfect Proxy Labels" Register: bit.ly/LSNI-2024 Info: networkinequality.com/lecture-series Complexity Science Hub

Starting in research is hard... So many papers, and they're written for experts! In a new initiative, Learning Theory Alliance is featuring ~monthly technical blog posts, highlighting recent results in an accessible manner. Follow us on Twitter to not miss out! 1st post 👇: 1/2

I am glad the ML community, across NeurICMLR and beyond, is starting to take this issue more seriously. If you participate in colluding behaviour, I very strongly recommend you stop *right now*. It's not worth ruining your entire career.

Congratulations to Justin and Pravesh!

The main reason I never wrote a multi-agent imitation learning paper before now is that I never actually understood what was *different* vs. the single-agent problem. Thanks to superstar Jingwu Tang's work, I am exceptionally excited to share we now do: arxiv.org/abs/2406.04219 1/

In [2406.01933] Orthogonal Causal Calibration (arxiv.org) we give a general reduction for calibration of causal models, e.g. CATEs. Can view it as calibration analogue of orthogonal statistical learning. Calibrate with orthogonal loss to achieve oracle calibration.

Happy to share our new work on "Strategyproof Decision-Making in Panel Data Settings and Beyond" at SIGMETRICS 2024, joint with Anish Agarwal, Chara Podimata, Steven Wu (1/8) dl.acm.org/doi/10.1145/36…

It took a lot of work over multiple years but I'm proud to be able to finally share my first #ICML2024 paper: a new, fundamental framework for designing efficient inverse RL algorithms that we're in the process of scaling to real robots: gokul.dev/hyper/! 1/

Congratulations to Professor Nicolas Christin (Nicolas Christin), CyLab faculty member and Co-Director of the Carnegie Mellon University Secure Blockchain Initiative, on being named the new department head at CMU's CMU Software and Societal Systems Department! Learn more: cylab.cmu.edu/news/2024/07/0…

Due to a series of lucky 7s 🎰 in terms of co-authors, I'll be co-presenting 8 papers at #ICML2024 this year: 4 at the main conference, 4 at the workshops! Please DM if you'd like to talk imitation / inverse RL, RLHF, causal RL, or just give me an excuse for an extra pastry!

1/ When is an algorithm an improvement over an existing (e.g., human) decision policy? Our #ICML2024 work offers an approach for recovering tighter relative performance intervals under unmeasured confounding by isolating comparison-related uncertainty. arxiv.org/abs/2404.00848

At the afternoon poster session today at #ICML2024 (Hall C 4-9, #2701), I will be presenting what is, in my opinion, one of the best papers I've written. In it, we introduce SPO (Self-Play Preference Optimization), a meta-algorithm for RLHF from diverse preferences. (1/n)