Congratulations to the winners of #AAAI2019 Student Poster and 3-minute presentations! #AAAI19 #AI

In our new work, we propose a framework for humans teaching robots to accomplish tasks using visual inputs: sites.google.com/corp/view/data… arxiv.org/abs/1909.12200

I am fortunate to have interned DeepMindAI where we extended GAIL to make it work better for robot manipulation tasks. Task-Relevant Adversarial Imitation Learning (TRAIL) robustly learns policies and rewards from pixels Paper arxiv.org/abs/1910.01077 Demo youtu.be/46rSpBY5p4E

this was one of my favourite projects, where I had to pretend to be bayesian to make visualization more "principled". it took so long for the paper to be formally accepted, all thanks to awesome Konrad Żołna and Krzysztof Geras ! sciencedirect.com/science/articl…

While classified-dependent saliency maps are useful for debugging, they are unsuitable for general-purpose visualizations and object localization. I, Krzysztof Geras snd Kyunghyun Cho describe how to get classifier-agnostic saliency maps in a recently published paper doi.org/10.1016/j.cviu….

We just released hundreds of hours of robot data and human feedback in the form of reward sketches for multiple tasks: github.com/deepmind/deepm… The agent we used is also open-sourced here: github.com/deepmind/acme/… For more information, see our project site: sites.google.com/view/data-driv…

I am glad to have taken part in this project! Our method, CRR, is very easy to implement and we have run it in multiple and diverse environments (robotic manipulation, control, locomotion). It always works!

RL Unplugged🔌: Offline RL benchmark that comes with both data and implementations of existing offline RL agents. It makes it easy to enter RL research and allows reproducibility.More agents in Acme will come soon. Github: git.io/JJUhd arxiv: arxiv.org/abs/2006.13888

How can RL be made usable in the real world? Offline RL is part of the solution but we need to pick hyperparameters using offline data too. Researchers show that certain simple approaches can be remarkably effective at offline hyperparameter selection: bit.ly/2EDpNvQ

Happy to share that our work on “Semi-supervised reward learning for offline reinforcement learning” is now available on arxiv! arxiv.org/abs/2012.06899

Gato🐈a scalable generalist agent that uses a single transformer with exactly the same weights to play Atari, follow text instructions, caption images, chat with people, control a real robot arm, and more: dpmd.ai/Gato Paper: dpmd.ai/Gato-paper 1/

Gather-Attend-Scatter (GATS), a novel module that combines pretrained foundation models operating at different rates into larger multimodal networks. Paper: arxiv.org/abs/2401.08525

Can I officially add game dev to my CV now? Proud to have collaborated with many brilliant minds behind Genie 🧞, a foundation world model trained exclusively from Internet videos!

Excited to see Imagen 3 now available to everyone! This was my final project at Google DeepMind, and I couldn’t have asked for a better way to close out my journey at GDM. I’m grateful to have worked alongside such incredibly talented and inspiring individuals. Thank you all!

let Konrad Żołna, Krzysztof Geras and me present you with a new algorithm for obtaining a saliency map extractor: arxiv.org/abs/1805.08249. unlike some of the recent approaches, we realized that such an algorithm needs to in fact try... arxiv.org/abs/1805.08249