Self-organized dynamics in plastic neural networks: bistability and coherence

Area of Science:

• Computational Neuroscience
• Neuroscience
• Artificial Intelligence

Background:

• Neural networks exhibit complex dynamics influenced by synaptic plasticity.
• Understanding the interplay between neural activity and synaptic changes is crucial for brain function and AI.
• Biologically realistic neuron models with synaptic potentiation and depression are key to studying network behavior.

Purpose of the Study:

• To investigate the combined dynamics of neural activity and synaptic efficiency in a fully connected network.
• To identify and characterize stable equilibrium states in networks with dynamic synaptic connections.
• To explore how external input statistics influence state transitions in plastic neural networks.

Main Methods:

• Utilized a mean-field technique to analyze equilibrium states of neural networks.
• Modeled biologically realistic neurons with simple synaptic plasticity (potentiation and depression).
• Investigated the impact of correlated and non-correlated external inputs on network states.

Main Results:

• Identified a class of bistable neural networks with two distinct stable equilibrium states.
• One state features strong connectivity and coherent responses; the other shows loose connectivity and non-coherent responses.
• Positively or negatively correlated external inputs drive transitions between these network states.

Conclusions:

• Plastic neural networks can switch between coherent and non-coherent states based on external input.
• The statistical properties of external input are critical for state transitions.
• This network property may provide insights into the mechanisms of unsupervised learning models.