Pre-Synaptic Pool Modification (PSPM): A supervised learning procedure for recurrent spiking neural networks.

Area of Science:

• Computational Neuroscience
• Spike-Based Computing
• Machine Learning

Background:

• Learning synaptic weights in spiking neural networks (SNNs) to match target spike trains is a key challenge.
• Supervised learning aims to optimize SNN weights by maximizing spike train similarity.

Purpose of the Study:

• To investigate if optimizing spike train similarity in recurrent SNNs leads to accurate weight discovery.
• To introduce and evaluate novel supervised learning rules for SNNs.

Main Methods:

• Proposed Pre-Synaptic Pool Modification (PSPM), a heuristic supervised learning rule using stochastic weight updates.
• PSPM requires no gradient information and makes minimal assumptions about neuronal properties or network structure.
• Evaluated PSPM's effectiveness on all-to-all SNNs and analyzed local vs. homeostatic weight update contributions.

Main Results:

• PSPM significantly improved spike train similarity for SNNs.
• Learned weights often diverged from the ground truth model weights, even with improved similarity.
• Spike train similarity was sensitive to local weight updates.
• Synaptic homeostasis contributed to learning other network activity measures, like avalanche distributions.

Conclusions:

• Optimizing for spike train similarity alone may not be sufficient for accurate connectome inference from spike data.
• Additional constraints on connectivity statistics might be necessary for precise connectome reconstruction.
• Local weight updates are important for spike timing, while homeostasis plays a role in broader network dynamics.