GLOW – The Global Latent Workspace: towards AI models of advanced cognition.
Recent advances have allowed artificial intelligence (AI) to achieve remarkable performances in many sensory, perceptual, linguistic or cognitive tasks. However, deep learning models still lack robustness and flexibility, as well as computational resources and training data. Our interdisciplinary project GLoW proposes to explore new cognitive architectures inspired by the brain, for a more robust, flexible and frugal cognition. It is based on the cognitive theory of the Global Workspace: a large-scale system integrating and distributing information between specialized modules (perception, language, decision, action), to give rise to more advanced forms of cognition. Our project will directly implement the global workspace (GLoW) in deep learning models of increasing complexity, and evaluate their correspondence with brain networks. It will provide an explicit evaluation of this fundamental neurocognitive theory, and push the limits of current systems to a new generation of AI.