An attention-based associative memristive spiking neural network and its application in unsupervised image classification

Unsupervised learning does not require manual annotation of training data, showing significant value in image classification applications of hardware systems. Unsupervised learning in the existing memristive spiking neural networks (MSNNs) mainly focuses on the synaptic adjustment process between anteroposterior neurons, limiting learning to local neural activity and neglecting internal connections between neural responses. Associative memory is a crucial way for the brain to achieve memory, which associates different stimuli through unsupervised learning to establish interconnected network memories. Meanwhile, the human visual system utilizes attention mechanisms to select important information from massive data, effectively reducing the number of input neurons and the scale of neural networks. This paper proposes a fully circuit-designed attention-based associative memristive spiking neural network (AAMSNN) and applies it to unsupervised image classification. Inspired by the attention mechanism, attention encoding circuits and attention selection circuits are designed to search for and select important information, reducing the number of input neurons of AAMSNN. The associative memristive spiking neural network module consists of Pavlovian associative memristive cross arrays, and achieves unsupervised image classification by adjusting the associative memory weight. The AAMSNN has smaller MSNN size and fewer memristors than other MSNNs, and achieves superior unsupervised image classification accuracy.