Member-only story
DeepMind’s New Super Model can Generalize Across Multiple Tasks on Different Domains
Gato is able to master tasks such as image classification, question answering or controlling a robotic arm.
Most deep learning models specialized on mastering a single task in a single domain. Recently, we have seen the emergence of multi-task models in a single domains such as transformers like GPT-3 in the language space. We have also seen models like OpenAI’s Dall-E that can combine knowledge from different domains to master a single task. However, the idea of having a single model mastering multiple tasks across heterogenous domains remains an elusive goals for deep learning. DeepMind is known for talking some of the toughest challenges in AI and, with a new research, they have set their eyes on neural network architectures that can generalize across tasks and domains.
In a recent paper, DeepMind introduced Gato, a transformer model that is able to perform multi-domain tasks such as playing Atari, stacking blocks or answer questions. These tasks are accomplished using the same network with the same weights.
