🔥GATO: the new SOTA from DeepMind
May 19, 2022 DeepMind published an article about a new generic universal agent outside the realm of text outputs. GATO operates as a multi-modal, multi-tasking, multi-variant universal policy. On the same network with the same weights, you can play Atari, record images, chat, manipulate blocks, and set other tasks depending on your context: generate text, determine optimal joint torques, detect points, etc.
GATO is trained on a large number of data sets including experienced agents in both simulated and sparse environments, in addition to many natural language and image data sets. At the GATO training stage, data from various tasks and modalities are serialized in a flat sequence of tokens, grouped and processed by a transformer neural network, similar to a large language model. Losses are masked so GATO only predicts action and text targets.
When the Gato is deployed, the demo prompt is tokenized, forming the initial sequence. The environment then emits the first observation, which is also tokenized and reproduced to the next one. GATO autoregressively selects an action vector one token at a time. Once all markers containing an action vector are selected (defined by environment specific actions), the action is decoded and placed in the environment, which performs the steps and produces a new observation. Then the procedure is repeated. The model always observes all observations and actions in its context window of 1024 tokens.
https://www.deepmind.com/publications/a-generalist-agent