Deep learning and the Global Workspace Theory

In recent years, deep learning has steadily improved the state of art in artificial intelligence (AI), but mainly for single, well-defined tasks or challenges. Novel advanced neural network architectures, possibly inspired by neuroscience, are needed to create more general-purpose AI systems with flexible and robust capabilities. The Global Workspace Theory (GWT), introduced over 30 years ago, proposed such an architecture; we now consider its implementation a deep-learning framework. global workspace was primarily designed account conscious information processing human brain, principle, associated functional advantages could generalize systems. turn, considering can help constrain neuroscientific investigations brain function consciousness. Recent advances have allowed (AI) reach near human-level performance many sensory, perceptual, linguistic, cognitive tasks. There is growing need, however, novel, brain-inspired architectures. (GWT) refers large-scale system integrating distributing among networks specialized modules higher-level forms cognition awareness. We argue that time ripe explicit implementations this theory using techniques. propose roadmap based on unsupervised translation between multiple latent spaces (neural trained distinct tasks, sensory inputs and/or modalities) unique, amodal Latent (GLW). Potential GLW reviewed, along implications. (Neuro) objects events interpreted according options they offer observer terms available uses (including mental usage) possible actions: their affordances. bottom-up top-down selection enter workspace, means matching query key vectors. automatic incoming from one selected module into format suitable space all other modules. (Neuro/AI) resulting simulation situations, without direct connection reality facts. objective two domains A B, whereby successive translations B back should retrieve original input. which flows external environment towards called discriminative, generative opposite direction; some be both (with bidirectional flow). how representations domain acquire meaning, associating them related (and unrelated) domains. contains internal copy each module’s space, used broadcast; recruiting amounts effectively connecting corresponding space. low-dimensional captures structure topology input output (for discriminative networks, respectively). system, operating independently GLW, capable it when (to achieve this, gets coupled workspace). machine algorithm networks. measure aims optimize via training. Crick Koch [47.Crick F. C. framework consciousness.Nat. Neurosci. 2003; 6: 119-126Crossref PubMed Scopus (935) Google Scholar], ensemble activity produced current state, yet not strictly part it. immediate subjective experience sensations, emotions, thoughts (etc.) phenomenal consciousness; access consciousness denotes reasoning executive control actions, including language. architecture deliberate planning reasoning, typically slow effortful compared perceptual awareness, well-practiced reflexive behaviors. application model problem problem. Domain adaptation subset transfer learning.