Date: Thursday, 10th April, 3:30 pm (Warsaw time)
Join online: zoom link
Goal-Directed Behaviour without Predictions: the Topological Alignment Hypothesis
Abstract
Goal-directed behaviour is characterized by an anticipation of future outcomes, by responsiveness to failure and success, and by flexibility when facing the same situational context. Dominant approaches in cognitive psychology and neuroscience explain these features by positing an internal (forward, generative) model in which action control is guided by predictions and prediction errors. In this contribution, we propose an alternative prediction-free mechanism in which the anticipation of action effects is sufficient for goal-directed control: the Topological Alignment hypothesis. This framework operates at three levels: functional, neural, and computational. At the functional level, it posits a multimodal topological space to bind together multi-sensory and motor events. This space aligns potential actions and their corresponding perceptual outcomes. Goals arise from this alignment and enable the selection of actions that yield desired effects. At the neural level, the framework draws on neuroscientific evidence showing that the cerebral cortex contains aligned topographical and topological maps. These maps form distinct action-perception domains within the fronto-parietal system. Each domain consists of a parietal area tightly connected to a corresponding motor/premotor area, providing biological support for the Topological Alignment hypothesis as a core principle of brain organization. At the computational level, a novel algorithm is presented to formalize this theory and to demonstrate the acquisition of a multimodal map of goals. The algorithm leverages topological dimensionality reduction to construct multimodal mappings. By aligning these mappings, the algorithm enables both a bottom-up, habitual as well as a top-down, goal-directed form of action control.