Embedding multiple trajectories in simulated recurrent neural networks in a self-organizing manner

Area of Science:

• Computational neuroscience
• Neural network modeling
• Synaptic plasticity

Background:

• Recurrent neural networks are vital for cortical computation due to complex dynamics.
• Understanding synaptic learning rules is key to stable neural trajectory formation.
• Existing models lack rules for embedding multiple, distinct neural pathways.

Purpose of the Study:

• Investigate unsupervised learning rules for recurrent networks.
• Achieve stable, sparse, and multi-trajectory neural activity.
• Develop methods to quantify network structure and dynamics.

Main Methods:

• Proposed and tested presynaptic-dependent scaling (PSD), an unsupervised learning rule.
• Developed a recurrence index to quantify network structure.
• Trained recurrent networks with single and multiple input patterns.
• Examined the combined effects of PSD and spike-timing-dependent plasticity (STDP).

Main Results:

• PSD successfully generated networks with stable, propagating activity, avoiding runaway excitation.
• Training with multiple patterns embedded multiple non-overlapping neural trajectories.
• Network recurrence increased with the number of training patterns.
• Parallel PSD and STDP enhanced trajectory incorporation and stability but shortened trajectory duration.

Conclusions:

• Presynaptic-dependent scaling is a viable learning rule for creating self-organizing recurrent networks with multiple embedded trajectories.
• The developed recurrence index quantifies network structural complexity.
• Findings offer insights into self-organizing principles for complex neural computation.