Homeostatic synaptic normalization optimizes learning in network models of neural population codes.

Area of Science:

• Computational Neuroscience
• Neural Coding
• Machine Learning

Background:

• Accurate statistical models are crucial for understanding neural population activity.
• Random Projection (RP) models offer accuracy, efficiency, and scalability.
• RP models can be implemented as biologically plausible shallow neural networks.

Purpose of the Study:

• To introduce a novel class of RP models learned by optimizing sparse projections.
• To evaluate the performance of these 'Reshaped RP' models compared to standard RP models.
• To investigate the role of biological features and synaptic normalization in model optimization.

Main Methods:

• Developed Reshaped RP models by optimizing sparse projections, mimicking synaptic connection changes.
• Incorporated biological features and synaptic normalization into the learning process.
• Compared Reshaped RP models against standard RP and fully connected neural networks using data from monkey cortical neurons.

Main Results:

• Reshaped RP models demonstrated superior accuracy and efficiency compared to standard RP models.
• Models incorporating biological features and synaptic normalization showed enhanced efficiency.
• The developed models exhibited homeostasis in firing rates and synaptic weights.
• Sparse homeostatic reshaped RP models outperformed fully connected neural network models.

Conclusions:

• Reshaped RP models provide a scalable, efficient, and highly accurate approach to population coding.
• Biological features, particularly synaptic normalization, optimize network performance and efficiency.
• Synaptic normalization plays a dual role in maintaining neural homeostasis and enhancing information encoding.