A Generalist Agent (Gato) - DeepMind's single model learns 600 tasks
Could you train one machine learning model to learn hundreds of tasks spanning text, computer vision, and playing videogames and controlling robots? In this video we go over DeepMind’s GATO that does this with a model that is simpler and smaller than you may think. It’s a GPT-like model that learns over 600 tasks. It opens the door to World Scope 4 as discussed in the Experience Grounds Language Video.
Contents:
Introduction (0:00)
Gato - A generalist agent paper (0:49)
Modalities of input and output data (2:36)
Running Gato as a control policy - Figure 3 (5:34)
Training datasets and tasks (6:47)
The Robotics task of stacking blocks - Figure 4 (7:06)
Results - so how well did Gato perform? (7:47)
The tokenization system - text and images (10:33)
Tokenization of discrete values like Atari controller state (12:10)
Packing sequences into timesteps and episodes (13:19)
Conclusion (14:06)
The paper: A Generalist Agent
https://www.deepmind.com/publications...
Authors: Scott Reed, Konrad Żołna, Emilio Parisotto, Sergio Gómez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Giménez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas
----
Twitter: / jayalammar
Blog: https://jalammar.github.io/
Mailing List: https://jayalammar.substack.com/
---
More videos by Jay:
Experience Grounds Language: Improving language models beyond the world of text
https://www.youtube.com/watch?v=WQm7-...
Language Processing with BERT: The 3 Minute Intro (Deep learning for NLP)
• Language Processing with BERT: The 3 ...
Watch video A Generalist Agent (Gato) - DeepMind's single model learns 600 tasks online without registration, duration hours minute second in high quality. This video was added by user Jay Alammar 05 September 2022, don't forget to share it with your friends and acquaintances, it has been viewed on our site 5,986 once and liked it 222 people.