Unsupervised spiking neural network based on liquid state machine and self-organizing map

The liquid state machine (LSM) is a reservoir computing paradigm and also a type of recurrent spiking neural network, combining the core strengths of both spiking neural networks (SNNs) and recurrent neural networks (RNNs). The architecture and functionality of the LSM are closely related to biological neural systems in many respects. Current research has extensively focused on the optimization of the structure and parameters within the liquid layer of LSM. However, there is a relative scarcity of studies on the readout layer, and its potential and effectiveness warrant further investigation. In this paper, a fully unsupervised spiking neural network approach is proposed, integrating the LSM with a spike self-organizing map (SOM). This integration results in a fully unsupervised SNN capable of performing visual clustering analysis on the readout layer of LSM and enhancing its biological interpretability. To rigorously assess the performance of our proposed method, we selected two well-established datasets for experimental validation: the MNIST dataset for image processing and the FSDD dataset for sound recognition. The experimental results indicate that our unsupervised learning approach achieved classification accuracies of 90.0 % and 88.0 % on these respective datasets. The model proposed in this paper utilizes randomly generated weights in the liquid layer, thereby bypassing the complex process of parameter optimization and simplifying the model's architecture. Additionally, the adoption of unsupervised learning eliminates the need for labeled data, significantly improving computational efficiency. These attributes provide our model with substantial practical utility and significance, particularly in environments where computational resources are limited and labeled data is not readily available.