Giovanni Granato

Overview

Topics:  My research focuses on the neurocognitive processes underlying human flexible cognition and behavior (e.g., executive functions). I proposed the concept of "goal-directed acquisition and manipulation of internal representations at the basis of flexible cognition", formalised by two neuro-computational theories: the “Motivated Categorical Perception theory” (Granato et al., 2022; Granato, 2022) and the "Three-components theory of flexible cognition" (Granato and Baldassarre, 2021, Granato et al., 2020). In short, I investigate how artificial/biological agents acquire (e.g., categorical perception) and manipulate (e.g., inner speech, Goal-directed Attention, etc.) their representations to flexibly achieve their goals. This research is extended to the investigation of goal-directed representations acquisition/manipulation at the basis of Consciousness and Metacognition. At last, these investigations show applications also in Machine Learning (Neuro-inspired ML, Generative Models, etc) and Robotics (e.g., Machine Consciousness).

Methods:  Overall, I adopts the integrated top-down / bottom-up method of “Computational Neuropsychology” (for a simplified comparison with "Computational Neuroscience" see our web site and this brief prospect: https://scitechconnect.elsevier.com/computational-neuropsychology-vs-computational-neuroscience/). In particular I develop simulated neuro-inspired embodied agents and compare their cognition/behaviour with that of humans. Moreover, the research line benefits from a high integration with Machine Learning methods (e.g., generative modelling, Recurrent Neural Networks, etc) and Neuro/Cognitive Robotics studies (e.g., Machine Consciousness).

Structure:  My research line shows a high systematzed and hierarchilca structure: Fields, Research Directions, Theoretical Proposals, Computational Models (see the figure below).  In particular, each research direction mainly involves a specific research fields :

• Representation learning/manipulation and Goal-directed Flexile Cognition ("Computational Neuropsychology")

• Representation manipulation and Inner-Speech ("Computational Psychiatry")

• Representation manipulation and Consciousness ("Consciousness science")

• Representation learning/manipulation in ML and Robotics ("Neuro/cognitive Robotics and ML")

However, "Computational Neuropsychology" and "Computational Psychiatry" are partially overlapping fields. Indeed, the second ("Representation manipulation and Inner-Speech") and third ("Representation manipulation and Consciousness ") research directions strongly interact with the first ("Representation learning/manipulation and Goal-directed Flexile Cognition").  Overall, these research directions leaded to many theoretical and computational contributions (see next section).

I carry out my research activity as research coordinator at the Laboratory of Embodied Natural and Artificial Intelligence (LENAI). For an overview of my research line in the context of my laboratory, see my lab's webpage: https://www.istc.cnr.it/en/group/lenai

research_plan_2023_2024.jpg

Theoretical and computational contributions

A high interdisciplinarity (integration of fields and directions) leaded to the development of three theories of cognition and related validated computational models.

Motivated Categorical Perception hypothesis (MCPH; Granato et al., 2022; Granato 2022):  the MCP hypothesis formalises the early motivational and sensory-motor processes leading to the emergence of adaptive perceptual representations (categorical perception). It proposes that the brain representation learning processes are based on the interactions between cortical and sub-cortical macro-systems (e.g. striate/exstrastriate cortices, basal ganglia, motor cortices), supporting specific computational functions (e.g. perceptual abstraction, motivational bias, motor selection). In particular, it proposes that (a) these macro-systems show different contributions of unsupervised learning (UL) and reinforcement learning (RL) along the sensory-motor hierarchy (UL in Striate cortex, UL + RL in Extrastriate cortices, RL in motor cortices) and (b) the emergences of a motivated categorical perception needs of motivated/goal-directed agent-environment continuous interactions (embodied perception). 

• Computational model and experimental validation: this hypothesis led to the development of a theory-based computational model, supported by a system-level neuro-inspired architecture, that performs a sorting task. We experimentally corroborated the hypothesis in Granato et al., 2022, where we compared different perceptual profiles of the model and perceptual functioning of autistic persons, thus explaining the altered categorical perception in Autism.

• References:

  • Granato G., Cartoni E, Da Rold F, Mattera A, Baldassarre G (2022) Integrating unsupervised and reinforcement learning in human categorical perception: A computational model. PLoS ONE 17(5): e0267838

  • Granato, G. (2022). Flexible goal-directed manipulation of representations: computational models of healthy and pathological human cognition (Doctoral dissertation, University of Plymouth).

mcp_hypothesis_website.png

Three-component theory (3CT; Granato and Baldassarre, 2021; Granato et al., 2020): the 3C theory formalises the neuro-cognitive processes that boost cognitive flexibility during goal-directed behaviours. Overall, it proposes that “flexible cognition depends on the integration between a top-down goal-directed manipulation of representations and sensory–motor interactions with the environment”. In particular, the theory proposes that three main elements are at the basis of flexible goal-directed behaviour: (a) Three key architectural components (executive Working Memory, Hierarchical Perceptual systems, Top-down manipulators), (b) First-order representations/manipulations (e.g., percepts - selective attention) and second-order representations/manipulations (e.g., goals - inner-speech) , (c) embodied sensory-motor loops.

• Computational model and experimental validation: this theory led to the development of two theory-based embodied computational models, supported by a system-level neuro-inspired architecture, that perform a rule-based sorting task (Wisconsin Card Sorting Test; Heaton, 2000). The second model expects the addition of an “inner-speech component”, that emulates a self-directed high-order representation manipulation. We experimentally corroborated the theory and models - thus reproducing the task performance - in Granato and Baldassarre (2021; young adults, old adults, frontal lesions, Parkinson),  Granato et al (2020; young adults in three different experimental conditions), Granato et al. (2022; autistic and neurotypical children, teenagers, young adults, middle adults), Granato et al. (2023; young adults, Schizophrenia). Overall, the model reproduced the cognitive and behavioural features of 17 human groups in healthy, clinical and neuro-divergent conditions.

• References:

  • Granato, G., & Baldassarre, G. (2021). Internal manipulation of perceptual representations in human flexible cognition: A computational model. Neural Networks, 143, 572-594.

  • Granato, G., Borghi, A. M., & Baldassarre, G. (2020). A computational model of language functions in flexible goal-directed behaviour. Scientific reports, 10(1), 1-13.

  • Granato, G., Borghi, A. M., Mattera, A., & Baldassarre, G. (2022). A computational model of inner speech supporting flexible goal-directed behaviour in Autism. Scientific reports, 12(1), 1-15.

  • Granato, G., Costanzo, R., Borghi A. M., Carruthers, S., Mattera, A., Rossell, S., & Baldassarre, G. (2024). Flexible Goal-directed Cognition and Inner-speech in Schizophrenia Spectrum Disorders: from Clinical Data to Computational Modeling, and Backward.  Comprehensive Psychiatry. "Under Revision". Pre-print: https://psyarxiv.com/n68yr

3ch_website_1.png

Goal-Aligning Representation Internal Manipulation (GARIM) theory (Granato and Baldassarre, 2023): the GARIM theory  extended the three-component theory, thus formalising neuro-cognitive processes at the basis of high-order cognition and Consciousness. The central idea of the GARIM theory is that “conscious states support the active manipulation of internal representations, making them more aligned with the goals pursued. This goal-oriented alignment leads to the generation of the necessary knowledge to face novel situations and goals and to make goal-directed behaviour more flexible and effective”. 

• Computational model and experimental validation: although the theory has not received any computational/experimental validation, it is compatible with the main theories of Consciousness and provides an architectural blueprint for future Computational Models and ML/Robotics systems.

• References:

  • Granato, G. & Baldassarre, G. (2023). A Neurocomputational Account of Flexible Goal-directed Cognition and Consciousness: The Goal-Aligning Representation Internal Manipulation theory (GARIM). Neural Networks. Under Review. Pre-print: arxiv.org/pdf/1912.13490.pdf

garim_website_1.png

Ongoing activities and future directions

I am working to extend my theoretical proposals and related computational models towards higher order cognition, including for example the fundamental concepts of Metacognition (Monitoring and Control). In accordance with this  direction, I will extend my theories to include "Visual Planning", thus further investigating the relationship between "representation manipulation" and higher order cognition (Planning/Problem Solving).

Furthermore, I will extend my investigations on Inner-Speech, thus comparing its different forms and alterations.

Finally, I expect to extend my theories and models to generate theoretical proposals and "proofs-of-concept" in Machine Learning and Robotics.