One Spiking Neuron Classification Based on Kolmogorov Complexity
This paper investigates the potential of a minimalist spiking neural network for digit recognition tasks, using the MNIST dataset as a benchmark. The proposed model features a single spiking neuron utilizing the Izhikevich neuronal model, deliberately crafted without weights or a learning phase, embodying the minimalist approach of maximizing performance with minimal resources. Our approach integrates self-organizing maps with a novel optimization method for cluster selection, leveraging Kolmogorov complexity and prototype abstraction. The model achieves 90% accuracy on inverted grayscale Arabic numerals. On Roman numerals, it maintains strong 87.4% accuracy, excelling particularly on well-separated patterns. These results confirms the model's suitability for regular inputs and highlights the potential of minimalist spiking architectures for future neuromorphic systems.