Lior Shani

Currently at Google Research.

PhD in Reinforcement Learning from the Technion.
Main research interests include Reinforcement Learning and Large Language Models

[Ofir Nabati @ofirnabati]

We're excited to share our new paper: "Personalized and Sequential Text-to-Image Generation"!
Check out the paper and our new sequential human rater dataset! 👇

Paper: https://t.co/kyvBHZ5Uja
Dataset: https://t.co/UmZ2RZJ3Ro

Details below..
1/N 🧵

[Lior Shani @LiorShan]

Check out our paper for more details on how we're pushing the boundaries of RLHF for more natural and effective multi-turn conversations!
https://t.co/RE9EDutUqr
@LiorShan @aviv_rosenberg @AsafCassel
(8/8)

[Lior Shani @LiorShan]

Excited to share our latest research on aligning Large Language Models (LLMs) with human preferences! We're moving beyond single-turn interactions to improve multi-turn conversations.
https://t.co/RE9EDutUqr

Joint work by @GoogleResearch @GoogleDeepMind
(1/8)

[Lior Shani @LiorShan]

Experimental results showing MTPO outperforms single-turn RLHF baselines and a multi-turn generalization of RLHF. This demonstrates the effectiveness of our approach in improving the quality of multi-turn conversations.
(7/8)

[Guy Tennenholtz @guytenn]

Check out our paper here: https://t.co/gPqkAmuRys
Code can be found here:
https://t.co/p8WjcLc434
And of course, thank you to everyone who collaborated with me on this project: @NadavMerlis, @LiorShan, @shiemannor, @ShalitUri, @GalChechik, Assaf Hallak, and @DalalGal.

(9/n)
n=9

[Lior Shani @LiorShan]

Finally, we discuss the need for exploration in imitation learning, and argue that our Apprenticeship Learning based approach which relies on the MDP structure, is superior to supervised learning approaches such as BC.

[Lior Shani @LiorShan]

Glad to present our paper "Online Apprenticeship Learning" at #AAAI2022
https://t.co/24tTXxl1bJ
with @TZahavy @MannorShie

We show how to efficiently reproduce experts' behavior from an offline data of trajectories, by interacting with the MDP (when rewards are not specified).

[Lior Shani @LiorShan]

We show our approach is both theoretically efficient and practical: we provide regret guarantees and show how to avoid solving an MDP at each iteration as in prior works. This allows us to devise a well-performing deep RL implementation of our (OAL) algorithm.

[Guy Tennenholtz @guytenn]

teleporting, swimming with sharks, and lots more. It turned out better than I've ever expected!

Now, before releasing it out to the world, I'm looking for a partner that can help me market the game properly. You can help me out by retweeting!

Promo:

[Ludwig Cancer @Ludwig_Cancer]

Ludwig @Princeton Director Joshua Rabinowitz, a pioneer of metabolomics, has contributed to the development of a cancer therapy & undone enduring assumptions about metabolism. His work is opening new approaches to cancer therapy. Learn more: https://t.co/yCUXGwtqpy

[Lior Shani @LiorShan]

Come chat today at the deep RL workshop @NeurIPSConf!

[Manan Tomar @manan_tomar]

Overall, MDPO is an easily scalable policy optimization algorithm with minimal hyper-params/heuristics involved, and is nicely grounded in mirror descent theory :)

Joint work with @LiorShan, Yonathan Efroni, Mohammad Ghavamzadeh

Come chat on Dec 11, 11:30 am PST!

[Lior Shani @LiorShan]

Applying Mirror Descent in deep RL is nice and easy!
https://t.co/Xky8VrV8u2

[Manan Tomar @manan_tomar]

Overall, MDPO is an easily scalable policy optimization algorithm with minimal hyper-params/heuristics involved, and is nicely grounded in mirror descent theory :)

Joint work with @LiorShan, Yonathan Efroni, Mohammad Ghavamzadeh

Come chat on Dec 11, 11:30 am PST!

[Lior Shani @LiorShan]

@manan_tomar https://t.co/Xky8VrV8u2

[Lior Shani @LiorShan]

Prof. Shie Mannor is presenting our work at the great RL theory seminar this Tuesday! The talk will be about the connections between TRPO and convex optimization, possible practical implications and how to explore in policy optimization...

[RL Theory Virtual Seminars @RLtheory]

Our next talk:
06/09: Shie Mannor (Technion)
"Adaptive Trust Region Policy Optimization: Global Convergence and Faster Rates for Regularized MDPs"

For details, please see the website:
https://t.co/zra0wITmwb

[Lior Shani @LiorShan]

Prof. Shie Mannor is presenting our work at the great RL theory seminar this Tuesday! The talk will be about the connections between TRPO and convex optimization, possible practical implications and how to explore in policy optimization...

[Gergely Neu @neu_rips]

our seminars are back this week with a real black-belt RL theorist!!