Internal simulation of perceptions and actions.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Neural networks are increasingly used to model complex cognitive functions.
• Developing architectures that can internally simulate perception and action is a key challenge.
• The Associative Self-Organizing Map (A-SOM) offers a novel approach to perceptual representation and association.

Purpose of the Study:

• To investigate neural network architectures capable of internally simulating perception and action.
• To evaluate the performance of architectures employing the Associative Self-Organizing Map (A-SOM).
• To assess the impact of recurrent connections and learning rules on network capabilities.

Main Methods:

• Development of multiple neural network architectures, including a bimodal perceptual architecture and others with action networks trained by the delta rule.
• Utilizing the Associative Self-Organizing Map (A-SOM) as the core perceptual component.
• Testing architectures with and without recurrent connections on simulation tasks.

Main Results:

• The bimodal perceptual architecture successfully simulated activity patterns without sensory input.
• An architecture without recurrent connections achieved 100% training and 80% test classification accuracy.
• A recurrently connected architecture maintained high accuracy (100% then 90%) for extended periods after input cessation.

Conclusions:

• The proposed A-SOM based neural network architectures show significant promise for simulating perception and action.
• Recurrent connections enhance the ability of these networks to maintain sequential outputs.
• The findings support the potential of these models for understanding and replicating cognitive processes.