DeepMind Unveils Gato, a Single AI Model Mastering 600 Diverse Tasks
DeepMind introduces Gato, a general-purpose transformer that tackles over 600 tasks including image captioning, robot manipulation, and Atari games using a single set of network weights. This multi-modal approach signals a significant step toward artificial general intelligence.
DeepMind introduces Gato, a single general-purpose transformer sequence model that operates like an AI Swiss Army knife. This innovative agent successfully performs over 600 diverse tasks, ranging from engaging in dialogue and captioning images to stacking blocks with a physical robot arm and navigating simulated 3D environments. Remarkably, Gato accomplishes all of these feats using the exact same network and weights, even outperforming human players in Atari games.
The researchers base their work on the hypothesis that training a generally capable agent on a massive number of tasks enables it to adapt to even more tasks with very little additional data. This general-purpose approach provides significant advantages over traditional methods, as it eliminates the need to hand-craft specific policy models for each individual field. Furthermore, it leverages a vast amount of diverse training data to drive continuous improvements across data, compute, and model scale.
To achieve this impressive flexibility, Gato processes massive multi-modal data by serializing it into a flat sequence of tokens. The model trains on a wide variety of datasets, including simulated and real-world agent experiences, the MassiveText dataset for language, and various vision-language datasets. By functioning as a multi-modal, multi-task, and multi-embodiment generalist policy model, Gato easily adapts to tasks with varying modalities and action specifications, representing a major step toward machine learning's ultimate goal of artificial general intelligence.