Cognitive Machinery and Behaviours

EasyChair Preprint 3609

Abstract

In this paper we propose to merge theories and principles explored in artificial intelligence and cognitive sciences into a reference architecture for human-level cognition or AGI. We describe a functional model of information processing systems inspired by several established theories: deep reinforcement learning mechanisms and grounded cognition theories from artificial intelligence research; dual-process theory from psychology; global-workspace theory, somatic markers hypothesis, and Hebbian theory from neurobiology; mind-body problem from philosophy. We use a formalism inspired by flow-graph and cybernetics representations. We called our proposed architecture IPSEL for Information Processing System with Emerging Logic. Its main assumption is on the emergence of a symbolic form of process from a connectionist activity guided by a self-generated evaluation signal. This theoretical work aims to provide a reference architecture for analysing behaviours of artificial systems. We also discuss artificial equivalents of concepts elaboration, common-sense and social interactions. The originality of this transdisciplinary work is that it stands at a general level of abstraction thus avoiding technical considerations. It can be considered as an artificial general intelligence design proposition althought its aim is to be an analysing tool for Human interactions with present and future artificial intelligence systems. We conclude by enunciating several conjectures on artificial agents’ behaviours which could allow readers to explore new perspectives on Artificial General Intelligence.

Keyphrases: Artificial Intelligence, Emerging logic, Formal model of General Intelligence, Reinforcement Learning, agent architecture, cognitive modeling

@booklet{EasyChair:3609,
  author    = {Bryan Fruchart and Benoit Le Blanc},
  title     = {Cognitive Machinery and Behaviours},
  howpublished = {EasyChair Preprint 3609},
  year      = {EasyChair, 2020}}