Space coding for sensorimotor transformations can emerge through unsupervised learning.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning in Neuroscience

Background:

• The posterior parietal cortex (PPC) is critical for integrating sensory information and posture signals into spatial reference frames for motor programming.
• Understanding the computational mechanisms underlying sensorimotor transformations in the PPC is essential for explaining goal-directed movements.

Purpose of the Study:

• To develop a computational model that mimics sensorimotor transformations within the PPC.
• To investigate if unsupervised learning can generate neural representations similar to those observed in the PPC for motor control.

Main Methods:

• A recurrent neural network (Restricted Boltzmann Machine) was trained using unsupervised learning on sensory data.
• The network learned a stochastic generative model of the sensory inputs.
• Motor programs were computed from hidden neuron activity via linear projection and delta rule learning.

Main Results:

• The model achieved an average motor error of less than 3°.
• Analysis revealed emergent gain-modulated visual receptive fields in the hidden neurons.
• These findings indicate that space coding for sensorimotor transformations can arise from unsupervised learning.

Conclusions:

• Unsupervised learning can effectively model sensorimotor transformations in a manner consistent with PPC function.
• Gain modulation emerges as an efficient coding strategy for integrating visual and postural information for motor command generation.