C-SOM SNN: A Continuous Self-Organizing Map Spiking Neural Network Based on STDP Unsupervised Learning Encoding with Designed Lateral Inhibition Distribution

To address traditional Spiking Neural Networks (SNNs) limitations in continuous variable encoding and biological plausibility, this study proposes C-SOM SNN, a novel SNN integrating SOM SNN clustering, CANN (Continuous Attractor Neural Network) continuous representation, and STDP unsupervised learning with good biological plausibility inspired by hippocampal spatial-coding neuron studies. Unlike conventional models confined to discrete variables, C-SOM SNN innovatively enables dual encoding of discrete and continuous variables via hybrid topology and Multi-Spike WTA (Winner-Take-All). Experiments validate its superiority on MNIST with fixed official train-test partition (five repeated runs with fixed seeds), and C-SOM SNN achieves 98.00 ± 0.18% accuracy versus 84.00 ± 0.40% (SOM-SNN) and 92.00 ± 0.31% (SNN-CANN). On Urban8K with official stratified 10-fold splits, it reaches 87.00 ± 0.65%, compared with 78.00 ± 0.92% (SOM-SNN) and 83.00 ± 0.84% (SNN-CANN), yielding measurable performance gains over baselines. In localization navigation, it stably encodes continuous landmark positions. It shows certain disjoint-cross-modal processing capability for visual and audio inputs in separate tasks by handling raw visual samples and audio signals transformed into image-like feature matrices. This work advances brain-inspired computing, offering an efficient, interpretable framework for complex tasks, with great potential for real-world disjoint-multi-modal applications.