Evolutionary learning in neural networks by heterosynaptic plasticity.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Training complex biophysical neuron models is crucial for understanding brain function.
• Traditional gradient-based methods like backpropagation struggle with instability and gradient issues in these models.

Purpose of the Study:

• To explore evolutionary algorithms (EAs) combined with heterosynaptic plasticity as a gradient-free training alternative.
• To develop a model inspired by biological mechanisms for training neural networks.

Main Methods:

• Utilized evolutionary algorithms (EAs) with a focus on heterosynaptic plasticity.
• Modeled agents with distinct neuron information routes, employing alternating gating and dopamine-driven plasticity.
• Incorporated biological mechanisms like dopamine function, dendritic spine meta-plasticity, memory replay, and cooperative synaptic plasticity.

Main Results:

• Neural networks trained with the proposed method exhibit brain-like dynamics during cognitive tasks.
• The method successfully trains both spiking and analog neural networks in feedforward and recurrent architectures.
• Achieved performance comparable to gradient-based methods on tasks such as MNIST classification and Atari games.

Conclusions:

• This research presents a robust, gradient-free alternative for training biophysical neuron models.
• The developed approach extends training methodologies, offering significant potential for advancing computational neuroscience and AI.