Multilevel development of cognitive abilities in an artificial neural network

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Artificial Intelligence

Background:

• Cognitive abilities emerge from postnatal environmental interactions.
• Neuronal mechanisms underlying cognitive development are complex and multifaceted.
• Existing models often lack a comprehensive, multi-level approach to information processing.

Purpose of the Study:

• To introduce a three-level computational model for information processing and cognitive ability acquisition.
• To define minimal architectural requirements for these processing levels.
• To investigate the influence of model parameters on performance and inter-level relationships.

Main Methods:

• Developed a three-level computational model: sensorimotor, cognitive, and conscious (Global Neuronal Workspace - GNW).
• Utilized visual classification, trace conditioning, and delay conditioning tasks to challenge model levels.
• Analyzed the necessity of epigenesis, dopamine, and interneurons for task performance.

Main Results:

• Epigenesis (synaptic selection/stabilization) is crucial for solving tasks at local and global scales.
• Dopamine is essential for credit assignment in delayed reward scenarios.
• Interneurons are necessary for sustained conscious representations in the GNW without sensory input.

Conclusions:

• The model demonstrates the importance of epigenesis, neuromodulation (dopamine), and specific neuronal populations (interneurons) in cognitive development.
• Balanced neural activity and excitatory/inhibitory ratios significantly impact model performance.
• The model offers insights into neurodevelopmental processes and potential AI architectures.