Development of Representations, Categories and Concepts—a Hypothesis

Abstract

A long-standing question in cognitive sciences and machine learning is how a system can develop high-level concepts and categories which are useful for motor and cognitive control. The author proposes an architecture which learns a hierarchy of increasingly abstract, invariant features. Invariance is achieved by selecting information which reflects distinctions present in supervisory signals conveyed by contextual inputs. The main hypothesis is that the right contextual information can be efficiently distributed by associations and attentional process. The original sources of contextual information are specialised systems which reflect the innate, hard-wired behavioural goals of the system. Sensorimotor coordination generates structured sensory stimuli and the intrinsic contextual signals can select the behaviourally significant structures