Optimal Localist and Distributed Coding of Spatiotemporal Spike Patterns Through STDP and Coincidence Detection

Area of Science:

• Computational Neuroscience
• Neural Coding
• Machine Learning

Background:

• Spiking neural networks process information through temporal patterns of neural activity.
• Understanding how individual neurons detect and encode specific spike patterns is crucial for deciphering neural computation.
• Previous theories primarily focused on localist coding, where one neuron represents one pattern.

Purpose of the Study:

• To extend existing theories to analyze how a single neuron can optimally detect multiple spatiotemporal spike patterns (distributed coding).
• To investigate the role of spike-timing-dependent plasticity (STDP) in enabling neurons to learn multiple patterns.
• To determine if neurons can achieve theoretical optimal signal-to-noise ratios (SNR) for pattern detection.

Main Methods:

• Analytical computation of the signal-to-noise ratio (SNR) for a multi-pattern-detector neuron using a threshold-free leaky integrate-and-fire (LIF) model.
• Simulation of a LIF neuron equipped with STDP, exposed to multiple input spike patterns embedded in Poisson spike trains.
• Investigation of the effect of STDP parameter tuning on pattern selectivity and coding efficiency.

Main Results:

• Theoretically, the SNR for detecting multiple patterns decreases slowly with an increasing number of patterns, remaining acceptable for tens of patterns.
• Simulations showed that STDP enables LIF neurons to become progressively selective for multiple input patterns without supervision.
• Optimal pattern detection was achieved by tuning STDP parameters and using a low adaptive threshold, allowing a single neuron to learn tens of patterns.

Conclusions:

• Coincidence detection and STDP are effective mechanisms for implementing distributed coding in single neurons.
• Neurons can learn to distinguish multiple specific spike patterns from random background activity.
• The findings support the compatibility of STDP with efficient distributed neural coding, applicable to feed-forward networks.