Iterative free-energy optimization for recurrent neural networks (INFERNO)

Area of Science:

• Computational Neuroscience
• Neuroscience
• Machine Learning

Background:

• The cortico-basal ganglia system forms a working memory crucial for planning neuronal sequences.
• Existing neurocomputational models struggle with long-range neural synchrony in spiking networks due to spontaneous activity.

Purpose of the Study:

• To propose a novel framework based on the free-energy principle to optimize neuronal sub-threshold activity for generating long neuronal chains.
• To address the challenge of controlling long-range neural synchrony in recurrent spiking networks.

Main Methods:

• Utilizing a stochastic gradient descent approach with a reinforcement signal (dopaminergic) to guide network activity.
• Employing an associative memory, modeled after the basal ganglia, to learn and control input vectors for the recurrent neural network.
• Framing spike synchrony as an optimization problem for sub-threshold neuronal activity.

Main Results:

• Demonstrated the capability to generate very long and precise spatio-temporal sequences (over 200 iterations) in habit learning and sequence retrieval tasks.
• Successfully applied the dual system to sequential planning of arm movements.
• Validated the model's ability to control recurrent neural networks for generating desired neuronal activity patterns.

Conclusions:

• The proposed framework effectively models the cortico-basal ganglia working memory for flexible, goal-directed neuronal sequence generation.
• Highlights the relevance of this approach for understanding novel architectures like Deep Networks, Neural Turing Machines, and the Free-Energy Principle.
• Offers a new perspective on controlling neural synchrony and generating complex spatio-temporal patterns.