Hosted by the School of Psychological Science
Abstract: Human cognition requires a flexible, general-purpose representation that rapidly combines arbitrary information in a usable form. I propose a simple neural architecture that can do this, the "plastic attractor". It can act as a low-capacity buffer, a pattern completer, and as a sequence encoder. I will examine three examples of humans' flexible representation. First, working memory binds arbitrary combinations of features into objects. Second, complex se can map complex stimulus to responses after a single instruction. Third, we can fill the roles of sentences with arbitrary contents. I show how plastic attractors can explain some classical features of working memory, attention and task sets, and can solve a major challenge in neural models of language: it can map the words in a sentence to their syntactic roles. I will show some cases in which the model agrees or disagrees with behavioural and neural data.