INCREMENTAL DEVELOPMENT OF A COGNITIVE ARCHITECTURE BASED ON THE GLOBAL WORKSPACE THEORY FRAMEWORK

Student thesis: Phd

Abstract

Different from mainstream AI research, Artificial General Intelligence emphasises the range of capabilities of the system and its ability to adapt across different domains. The gap between state-of-the-art and the ultimate goal of general intelligence is recognised. To fill the gap, this thesis focuses on the approach of cognitive architectures. The thesis aims to develop a cognitive architecture for robotic agents based on the Global Workspace Theory of human consciousness. The research focuses on the design of architecture within which various modules according to human brain functionalities interact with each other and function as a whole system to demonstrate intelligence behaviours based on the consciousness mechanism. With the cognitive framework defined in Global Workspace Theory, three aspects, attention, episodic memory and continual learning of human intelligence are studied and implemented in this thesis. To validate the development of the architecture, experiments are conducted to examine the agent’s abilities related to these three cognitive features. With attention mechanisms integrated into the Global Workspace framework, the agent demonstrates shifts of attention towards different stimuli with the combined ef- fect of top-down and bottom-up attention. In addition, the agent replicates the attentional blink and lag-1 sparing effect, two cognitive phenomena of human cognition. For episodic memory, the working definition is given, including three aspects and three cognitive processes of episodic memory. With the interaction between episodic memory and the consciousness mechanism defined in Global Workspace Theory, the experiment shows that the agent can not only fulfil inference tasks based on memory but also demonstrate the dynamics of memory curves. Last, with the implemented attention and episodic memory mechanisms, the robotic agent learns object-name pairs incrementally during the interactions with the human peer based on newly implemented mechanisms including self-knowledge, selective memory replay and memory consolidation. The novelties are summarised. By implementing different cognitive correlates, this thesis strengthens the feasibility of the Global Workspace framework for cognitive architecture development. This would bring more researchers’ attention to the employment of this framework for system development in the field of cognitive robotics. The study of attention mechanisms first time explains both attentional blink and lag-1 sparing effects in a computational way. The implementation of episodic memory summarises the literature and provides a working definition for computational development in cognitive architecture. As for the continual learning scheme, a self-knowledge-based selective memory replay approach is proposed to make the agent autonomously decide the importance of the learning information, saving memory consumption during learning. In addition, a general framework combining the Global Workspace Theory and Semantic Pointer Architecture is proposed for long-term development. In the discussion and conclusion, limitations of this thesis on cognitive architecture development are recognised, with corresponding solutions proposed for future work.
Date of Award1 Aug 2024
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorAntonio Chella (Supervisor) & Angelo Cangelosi (Supervisor)

Keywords

  • Global Workspace Theory
  • Machine Consciousness
  • Attention
  • Episodic Memory
  • Continual Learning
  • Cognitive Architecture

Cite this

'