🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Most people don't know that although most of my research work is in RL, I spent a significant portion of my PhD & early career on generative modelling (text, images, data compression). Building new RL algorithms for LLM training is a real delight - putting two passions together.

It took us 2+ years to figure out exactly how to think about, & work with a distributional version of the successor representation - doubly proud of this work by Jesse Farebrother and Harley Wiltzer that both lays down a mathematical foundation and improves on γ-models! Also, A+ visuals.

On the back of our 2017 distributional RL paper Martha White and Ehsan Imani wrote a piece showing that you can do regression better with a classification loss... that seemed wild at the time, but Jesse Farebrother, Rishabh Agarwal and co pushed this further and the results are amazing!

Love to build ML software? Reliant AI is hiring in Montreal. apply.workable.com/reliant-ai/j/5…

Amazing piece of work by Jesse Farebrother , Rishabh Agarwal , & star co-authors digging into classification losses in RL and their unreasonable effectiveness in a problem space that has mostly been dominated by regression methods. Don't miss this talk at ICML!

Because one level of distributions isn't enough - don't miss tomorrow's ICML spotlight by Harley Wiltzer , Jesse Farebrother , Arthur Gretton , and Mark Rowland: lifting the successor representation to distributions and moving the needle on what you can do with technique like γ-models.

Distributional successor features: A follow up to our distributional successor representation by my students Harley Wiltzer and Jesse Farebrother - those manim animations are quite something!

Interested in using reinforcement learning to train LLMs for problems where there’s no room for error? Do you want to build massive data pipelines to transform how we interact with scientific knowledge? We're hiring for multiple roles at Reliant: apply.workable.com/reliant-ai

🚀 Extremely excited about our latest work on Distributional RL algorithms for *high-frequency control*, to be presented at #neurips2024! Incredible collaboration with the OT wizard Yash Jhaveri, Marc G. Bellemare, David Meger, Patrick Shafto. Paper: arxiv.org/pdf/2410.11022

we've used Atari games as an RL benchmark for so long, but for a little while it's bugged me that it's a discrete action problem, since the original joysticks were analog... Jesse Farebrother & i fix this by introducing the Continuous ALE (CALE)! read thread for details! 1/9

We're hiring at Reliant AI! On engage chez Reliant AI! If you or someone you know is excited to build the future of AI-powered research - take a look, share widely. Montreal and Berlin (platform eng., ML eng., and RS) and North America (commercial). apply.workable.com/reliant-ai/

Take a look at this amazing piece of work by my student Jesse Farebrother - a new kind of world model based on successor representations that's a lot more robust than prior iterations. Incredible to see all the progress we've made in the last 5 years in RL.

Goodbye Toronto! So many serendipitous meetings Toronto Tech Week, incredible energy. Learned that Isaac Souweine and I like the same parties. Met too many AI founders to count, all making amazing new things. Now back to building!

A can-of-worms type question to all of you, from an RL researcher turned NLP startup founder: assuming a benchmark with no annotator error, is 100% accuracy on question answering possible?

Heading to Vancouver for #ICML2025 to present our work: Temporal Difference Flows. Make sure to check out the oral to learn how we’re now able to scale this exciting world model framework based on the successor representation! Also, feel free to reach out to discuss anything RL!

AI isn’t about saving time – it’s about doing what you couldn’t do before. If you could read ten thousand papers a day, you would ...