Clare Lyle (@clarelyle) / X

Clare Lyle

213 posts

Clare Lyle

@clarelyle

RL researcher who sometimes blogs at facilisdescensus.substack.com. Formerly @OATML_Oxford, now DeepMind.

Oxford

Joined October 2012

Clare Lyle
@clarelyle
Mar 17, 2025
📣📣 My team at Google DeepMind is hiring a student researcher for summer/fall 2025 in Seattle! If you're a PhD student interested in getting deep RL to (finally) work reliably in interesting domains, apply at the link below and reach out to me via email so I know you aplied👇
83K
Clare Lyle
@clarelyle
Dec 1, 2020
**coming soon to a NeurIPS near you** Do models that train faster also generalize better? We* apply tools from Bayesian probability theory to find out! arxiv.org/abs/2010.14499 * @miouantoinette @Robin_Ru @yaringal @markvanderwilk (1/9)
arxiv.org
A Bayesian Perspective on Training Speed and Model Selection
We take a Bayesian perspective to illustrate a connection between training speed and the marginal likelihood in linear models. This provides two major insights: first, that a measure of a model's...
Clare Lyle
@clarelyle
Jul 21, 2022
At #ICML today: why is generalization so hard in value-based RL? We show that the TD targets used in value-based RL evolve in a structured way, and that this encourages neural networks to ‘memorize’ the value function. 📺 icml.cc/virtual/2022/p… 📜 proceedings.mlr.press/v162/lyle22a.h…
Clare Lyle
@clarelyle
Apr 27, 2022
Today at #ICLR22 : Deep RL agents have to fit a series of value functions -- we show that this can make neural networks **worse** at fitting new targets later in training, and explore the implications of this in deep RL. 📜arxiv.org/abs/2204.09560 📺iclr.cc/virtual/2022/p…
Clare Lyle
@clarelyle
Jul 14, 2020
Excited to present our #ICML2020 paper on generalization to new environments in RL using tools from causal inference! Poster sessions today at 11pm BST and tomorrow at 12pm. Paper: arxiv.org/abs/2003.06016 ICML poster: icml.cc/virtual/2020/p…
arxiv.org
Invariant Causal Prediction for Block MDPs
Generalization across environments is critical to the successful application of reinforcement learning algorithms to real-world challenges. In this paper, we consider the problem of learning...
Clare Lyle
@clarelyle
Apr 9, 2021
Everyone knows auxiliary tasks improve RL agents' representations… but what does “improving the representation” actually mean? We (Mark Rowland* @wwdabney @georgostrovski) look to learning dynamics to answer this question in our upcoming AISTATS paper! arxiv.org/abs/2102.13089
Clare Lyle
@clarelyle
Feb 1, 2019
I spent this summer trying to figure out why distributional RL works, with @marcgbellemare and @pcastr. Our quest turned into a #AAAI19 paper that's now up on @arxiv_org arxiv.org/pdf/1901.11084…
Clare Lyle
@clarelyle
Aug 18, 2020
Q-networks HATE this ONE WEIRD TRICK to get near-optimal performance on cartpole-v0 in @OpenAI gym: a = 1 if obs[2] + obs[3] > 0 else 0 (for real tho this is a super convenient behaviour policy for debugging your deep RL model)
Clare Lyle
@clarelyle
Jul 23, 2023
Can’t wait to present our work on plasticity in neural networks at #ICML2023 this year! Reach out if you want to chat about optimization challenges in reinforcement / continual learning
Google DeepMind
@GoogleDeepMind
Jul 20, 2023
Going to #ICML2023? We’ll be sharing our latest advances in AI, covering themes such as: 🌐 AI in the (simulated) world 💡 The future of reinforcement learning ⭕ Challenges at the frontier of AI Find out more now: dpmd.ai/44Td7I9
29K
Clare Lyle
@clarelyle
Dec 6, 2021
“Train faster, generalize better” for NAS: our @NeurIPSConf spotlight proposes a simple measure of training speed achieving superior performance estimation in a range of NAS settings. @FilMiroslav @ox_robin @miouantoinette @markvanderwilk @yaringal (1/4)
arxiv.org
Speedy Performance Estimation for Neural Architecture Search
Reliable yet efficient evaluation of generalisation performance of a proposed architecture is crucial to the success of neural architecture search (NAS). Traditional approaches face a variety of...
Clare Lyle
@clarelyle
Apr 19, 2019
Why does choosing your prior after training break PAC Bayes bounds? I couldn't find a concise answer online so I wrote one up here: clarelyle.com/posts/2019-04-…
Clare Lyle
@clarelyle
Sep 13, 2019
Twitter-verse! Know of any social impact organizations that could benefit from a team of Oxford grad students working with them on ML solutions to meaningful problems? RAIL is looking for project partners for our next cycle- more info here:
rhodeslab.com
Rhodes Artificial Intelligence Lab (RAIL)
AI to Improve the World Learn More
Clare Lyle
@clarelyle
Aug 5, 2023
Looking forward to speaking at CoLLAs this year and making my first pilgrimage back to McGill as a non-student! 😊
CoLLAs 2026
@CoLLAs_Conf
Aug 4, 2023
Did you know that August is prime time for stargazing in Montreal? But why look to the sky when you can spot rising star Clare Lyle (@clarelyle) at #CoLLAs2023? 🌟 Join us for a celestial spectacle of machine learning brilliance! Register now: lifelong-ml.cc
9.4K
Clare Lyle
@clarelyle
Oct 1, 2020
Looking forward to presenting at today's session (and feeling a little starstruck to be on the same lineup as these folks 🤩)!
Marc G. Bellemare
@marcgbellemare
Oct 1, 2020
Looking forward to today's Deep RL Theory Workshop at the virtual @SimonsInstitute - from language to latent states! With @jacobandreas @yayitsamyzhang @clarelyle @ShamKakade6 and Doina Precup, hosted by none other than @LihongLi20 . See you there! simons.berkeley.edu/workshops/sche…