A neural network model for the acquisition of a spatial body scheme through sensorimotor interaction

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Developing internal representations of 3D space is crucial for autonomous systems.
• Current methods often require labeled data or predefined sensory mappings.

Purpose of the Study:

• To present a novel unsupervised sensory matching learning technique.
• To develop an internal representation of three-dimensional (3D) information invariant to sensory modalities.
• To demonstrate the acquisition of this representation through sensorimotor interactions.

Main Methods:

• Utilized a neural network model simulating a creature's sensorimotor system.
• Employed a tactile-sensitive body and a multi-degree-of-freedom arm with proprioceptive feedback.
• Trained the model using unsupervised learning based on sensory-motor experiences.

Main Results:

• Successfully acquired an internal 3D representation invariant to sensory input.
• Demonstrated the development of a distributed body scheme representation.
• Achieved convergence of learning through computer simulations with a 7-DoF arm and 20 tactile fields.

Conclusions:

• Unsupervised sensory matching learning can effectively build 3D representations.
• Sensorimotor interactions are sufficient for acquiring complex internal representations.
• The proposed technique offers a pathway for more adaptable and generalizable AI systems.