Unsupervised Spiking Neural Network with Dynamic Learning of Inhibitory Neurons

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) mimic the human brain's processing of timing information using discrete spikes.
• Conventional SNNs utilize models like leaky integrate-and-fire, spike timing-dependent plasticity, and adaptive thresholds.

Purpose of the Study:

• To propose a novel, biologically plausible Spiking Neural Network (SNN) for image recognition tasks.
• To enhance SNN performance by introducing new biological models for dynamic inhibition, synaptic wiring, and Bayesian inference.

Main Methods:

• The proposed SNN integrates biologically plausible paradigms with novel components: dynamic inhibition weight change, a Hebbian-based synaptic wiring method, and Bayesian inference.
• Unsupervised learning is employed, with dynamically changing inhibition weights influencing synaptic wiring and neuronal populations.
• Bayesian inference is used in the network's inference phase for digit classification via spike counting.

Main Results:

• The proposed biologically plausible SNN model demonstrates improved performance in image recognition tasks compared to existing SNN models.
• The integration of dynamic inhibition, synaptic wiring, and Bayesian inference contributes to enhanced classification accuracy.

Conclusions:

• The novel SNN architecture offers a more bio-realistic approach to artificial neural networks.
• The proposed model represents a significant advancement in biologically plausible SNNs for image recognition, outperforming previous models.