Force Learning in Balanced Cortical E-I Networks

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Neural Networks

Background:

• Force learning is a training method for recurrent neural networks (RNNs), a type of artificial neural network, related to reservoir computing (RC).
• While effective in machine learning, the brain's capacity for force learning remains underexplored.
• Reservoir computing typically uses fixed random weights, whereas force learning trains these synaptic weights.

Purpose of the Study:

• To investigate the feasibility of force learning in biological neural networks.
• To model the cerebral cortex using an excitatory-inhibitory (E-I) network and apply force learning.
• To determine the conditions under which force learning is most effective in an E-I network.

Main Methods:

• An excitatory-inhibitory (E-I) network, modeling the cerebral cortex, was constructed with multiple modules.
• A readout mechanism calculated network output based on filtered firing rates of excitatory neurons.
• Feedback connections were introduced, and the network was trained using force learning to generate sinusoidal signals.

Main Results:

• The E-I network exhibited transitive chaotic synchronization.
• Force learning efficiency was maximized at an optimal E-I balance, near the edge of chaos.
• The study demonstrated that force learning effectiveness is enhanced by the interplay between excitatory and inhibitory neurons.

Conclusions:

• The cooperation between excitatory and inhibitory neurons is crucial for effective force learning in neural networks.
• This finding suggests a potential mechanism for complex dynamics generation in the brain through E-I network interactions.
• Force learning in E-I networks offers insights into brain computation and learning.