Block-Clustering on Neural Networks for Large-scale Memristor-based Implementation

The emergence of memristors with tunable resistance states provides building blocks for neuromorphic computing systems (NCSs) as electric synapses. However, the large-scale realization of memristive neural networks confronts with challenges, that is, a rapid increase in network complexity causes high-cost mapping and low-reliability of memristor connections. The general method is to make neural networks sparse. But most sparse networks actually reduce the utilization rate of the memristors and the expected large scale is normally beyond that can be offered by current integration technology of memristors. In this paper, a reliable block-clustering framework (BCF) with universality is proposed for both directed and undirected memristive neural networks. The idea of sparse and block clustering can realize the use of fewer memristor integrated circuits, and solve the current problem of difficult large-scale integration of memristive arrays. BCF integrates block clustering approaches and memristor array properties to cluster original large sparse synapse connections into several smaller and denser groups as much as possible. Then, the denser connections can be realized by memristor crossbar arrays with enhanced utilization rate and the small number of intra-group connections can be realized by discrete memristor synapses. BCF consists of iterative spectral clustering (ISC), swap clustering (SC) and hybrid ISC+SC, where these three methods are evaluated according to the criteria including outlier percentage, time consumption. Experiment results and analysis demonstrate that BCF can be effectively used on neural networks with different sparsities. This novel clustering framework provides a promising key technique for improving the utilization of memristor crossbar array and guaranteeing the communication quality of the whole system, which is conductive to promote large scale NCSs implementation.