Learning in Wilson-Cowan Model for Metapopulation

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• The Wilson-Cowan model is a foundational neural mass network model simulating brain region dynamics.
• Metapopulation models extend this by connecting multiple neural regions, representing complex brain networks.
• Existing models often lack mechanisms for stable memory or learning.

Purpose of the Study:

• To integrate stable attractors into the Wilson-Cowan metapopulation model.
• To transform this enhanced neural mass network into a biologically inspired learning algorithm.
• To evaluate the algorithm's performance on various benchmark classification tasks.

Main Methods:

• Incorporated stable attractor dynamics into the Wilson-Cowan metapopulation framework.
• Developed a novel biologically inspired learning algorithm based on the modified model.
• Tested the algorithm's classification accuracy using datasets like MNIST, Fashion MNIST, CIFAR-10, TF-FLOWERS, and IMDB.
• Combined the algorithm with convolutional neural networks and transformer architectures (BERT).

Main Results:

• The enhanced Wilson-Cowan metapopulation model successfully learned and performed classification tasks.
• High classification accuracy was consistently achieved across diverse datasets (MNIST, Fashion MNIST, CIFAR-10, TF-FLOWERS, IMDB).
• The model demonstrated robust performance when integrated with both CNNs and BERT architectures.

Conclusions:

• Stable attractors can be effectively incorporated into metapopulation neural mass models.
• This modification transforms the model into a powerful, biologically inspired learning algorithm.
• The approach shows significant promise for advancing machine learning and computational neuroscience applications.