Internal manipulation of perceptual representations in human flexible cognition: A computational model

Abstract

Executive functions represent a wide set of goal-directed cognitive processes that rely on integrated cortical-basal ganglia brain systems and are at the basis of human flexible behaviour. Several computational models have been proposed to study cognitive flexibility and the Wisconsin Card Sorting Test (WCST), an important neuropsychological test used for measuring such function. These models clarify important aspects of cognitive flexibility, in particular the processes concerning decision making, motor response, and feedback-dependent learning. However, several studies suggest that categorisation processes, as those involved in the solution of the WCST, also involve a fundamental category representation stage supporting the other processes. Surprisingly, all models of the WCST ignore such fundamental stage and propose that decision making directly triggers actions. Here we propose a novel hypothesis for which the key element of cognitive flexibility resides on the acquisition of suitable representations of percepts, and their top-down internal manipulation, to prepare effective actions. We also propose a neuro-inspired computational model to operationalise the hypothesis. The model is validated by systematically reproducing and interpreting the behaviour of healthy young and old adults, and of frontal and Parkinson pathological patients. The hypothesis and model also allow the proposal of a new version of the WCST that might be used to further investigate the important role of the internal manipulation of representations here proposed to be at the core of flexible goal-directed behaviour.

Integrating unsupervised and reinforcement learning in human categorical perception: A computational model

“…For example, during the solution of a categorisation task sensory processes increase the between-category differences and decrease the within-category difference, a phenomenon called ‘categorical perception’ (CP; [ 1 , 2 ]). In [ 3 ] we corroborated the idea that, during category learning, a CP-like effect can be supported by a top-down selection of perceptual representations. In that work, we have assumed that perceptual learning processes have previously created the category-based representations, which the agent select during the task solution.…”

“…The model does not use a true ‘learning process’, in that the emergence of categorical perception is based on a top-down inferential process. In this sense, the model has common features with our previous works on representation manipulation [ 3 , 37 ], in which a recurrent neural network biases a sensory system and leads to the emergence of categorical perception. Moreover, the inferential process is not influenced by a performance-related feedback signal.…”

“…This section presents the description of the task and of the computational model. Note that, as previously done [ 3 , 37 ], a neuroscientific and theoretical investigation of brain systems supporting the investigated phenomenon (categorical perception) preceded and guided the model building. Therefore, in section 2.2 we describe the theoretical and neural underpinnings of the model functioning, while in section 2.3 we describe the computational algorithms and artificial neural networks that support the model components.…”

“…As in [ 3 ], the model abstracts the fine-grain biological details (e.g. neuronal micro-circuitry or bio-grounded plasticity).…”

“…Reconciling controversial results [ 13 ], propose that perceptual learning processes occur at different stages of visual hierarchy, depending on the task demands. In addition, several studies [ 3 , 14 , 15 ] suggest that the subcortical structures that support reward-based feedback signals (e.g. basal ganglia) interact with cortical structures, contributing to the emergence of category-based perception.…”