Artificial Neural Network (ANN) to Spiking Neural Network (SNN) Converters Based on Diffusive Memristors. (29th March 2019)

• Record Type: Journal Article
• Authors: Midya, Rivu; Wang, Zhongrui; Asapu, Shiva; Joshi, Saumil; Li, Yunning; Zhuo, Ye; Song, Wenhao; Jiang, Hao; Upadhay, Navnidhi; Rao, Mingyi; Lin, Peng; Li, Can; Xia, Qiangfei; Yang, J. Joshua
• Is Part Of: Advanced Electronic Materials. Volume 5:Number 9(2019)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2019-03-29
• Subjects: diffusive memristor -- memristor synapse -- oscillatory neuron; Materials -- Electric properties -- Periodicals; Materials science -- Periodicals; Magnetic materials -- Periodicals; Electronic apparatus and appliances -- Periodicals; 537
• DOI: 10.1002/aelm.201900060
• Languages: English
• ISSNs: 2199-160X

Abstract

Abstract: Biorealistic spiking neural networks (SNN) are believed to hold promise for further energy improvement over artificial neural networks (ANNs). However, it is difficult to implement SNNs in hardware, in particular the complicated algorithms that ANNs can handle with ease. Thus, it is natural to look for a middle path by combining the advantages of these two types of networks and consolidating them using an ANN–SNN converter. A proof‐of‐concept study of this idea is performed by experimentally demonstrating such a converter using diffusive memristor neurons coupled with a 32×1 1‐transistor 1‐memristor (1T1R) synapse array of drift memristors. It is experimentally verified that the weighted sum output of the memristor synapse array can be readily converted into the frequency of oscillation of an oscillatory neuron based on a SiO x N y :Ag diffusive memristor. Two converters are then connected capacitively to demonstrate the synchronization capability of this network. The compact oscillatory neuron comprises multiple transistors and has much better scalability than a complimentary metal oxide semiconductor (CMOS) integrate and fire neuron. It paves the way for emulating half center oscillators in central pattern generators of the central nervous system. Abstract : Artificial neural networks (ANNs) can perform increasingly complicated functions. However, spiking neural networks (SNNs) are behind ANNs in performing similar tasks. Consolidating the advantages of these networks through an ANN–SNN converter is more advantageous. A proof‐of‐concept study of this idea is demonstrated experimentally by implementing hardware neurons coupled with a 32×1 1‐transistor 1‐memristor (1T1R) synapse array.