Fully hardware-implemented neuromorphic systems using TaO -based memristors
Jin Tian, Kang Lv, Jianzhong Chen, Le Zhang, Xin Guo
Abstract
Ongoing AI advancements confront challenges in physical limitations and architectural inefficiencies of the von Neumann architecture, prompting a shift toward neuromorphic designs inspired by biological brains. We present a hardware-based neuromorphic system leveraging TaO x -based memristors, facilitating regulation of synaptic arrays and development of a convolutional spiking neural network (CSNN) for offline learning-based tasks. The CSNN system incorporates algorithms including the spike-rate-dependent plasticity (SRDP) and the Manhattan rule to foster cooperative and competitive relationships among neurons. The integration of gating switches within neurons emulates the functionality of ion channels in biological neurons, enabling spatiotemporal integration. We also introduce methodologies for optimizing the integration of the memristor technology, addressing existing challenges such as non-ideal characteristics of memristors.