Adaptive training of cortical feature maps for a robot sensorimotor controller.

Area of Science:

• Computational Neuroscience
• Robotics
• Artificial Intelligence

Background:

• Traditional Self-Organising Maps (SOMs) require modifications for autonomous robotic applications.
• Existing SOM training methods often rely on predefined datasets and parameter reduction schedules.
• There is a need for self-regulating and adaptable feature maps in robotics.

Purpose of the Study:

• To develop a self-organising cortical feature map (SOFM) training system for autonomous humanoid robots.
• To create a self-regulating training process that does not require predefined training files or parameter reduction schedules.
• To enhance map flexibility for accommodating new information while preserving learned patterns.

Main Methods:

• Implementation of SOFMs using spiking neural networks.
• Development of a 'plasticity resource' (PR) as a global parameter to control map development rate.
• Utilisation of random input generation from exemplar patterns instead of predefined datasets.

Main Results:

• The plasticity resource (PR) effectively controls map training, replacing traditional learning rate parameters.
• Random input generation allows training without pre-specifying data volume.
• The PR's ability to increase or decrease enables maps to adapt to changing input conditions.

Conclusions:

• The developed 'Adaptive Plasticity' approach offers a self-regulating and flexible training method for SOFMs.
• This system addresses limitations of traditional SOMs in autonomous robotic applications.
• The PR mechanism facilitates continuous learning and adaptation in neural maps for robotics.