Neuromorphic pattern learning using HBM electronic synapse with excitatory and inhibitory plasticity

Abstract

Bio-inspired neuromorphic system has become a popular domain of research because of its promising potential for low-power and robust fault-tolerant computing beyond the contemporary Von Neumann architecture [1]. The elementary building block of artificial neuromorphic systems is typically depicted as synaptic devices connecting between pre- and post-neuron units (Fig. 1), mimicking the morphology of synapses and neurons in biological systems, e.g. in the human brain [2]. Generally, the connecting strength of synapse is called synaptic weight, which is plastic and can be adjusted by applying an appropriate learning rule, such as the winner-take-all rule [2-3], through the fired signals of pre- and post-neurons. The strengthening (excitatory)/ weakening (inhibitory) processes of synaptic weight are referred as potentiation (P) and depression (D). Furthermore, the rapid development of resistive-switching random access memory (RRAM) recently has inspired significant interests on its memristive applications as high-density electronic synapses in artificial neuromorphic systems [4]. However, most RRAM devices reported in the literatures can only perform gradual SET or gradual RESET operations, and cannot be used as excitatory and inhibitory synapses simultaneously [3, 5-6]. In this paper, we report on a homogeneous barrier modulation (HBM) RRAM [7] that is capable of a simultaneous P and D (P+D) operational scheme. We perform a simulation of pattern learning algorithm based on the winner-take-all rule and experimental synaptic characteristics. The P+D scheme improves the contrast development of pattern learning and immunity to input noise as compared with the P-only scheme. The tolerance on the variations of synaptic cells is also examined with randomness at the initial resistance and P/D characteristics. This study suggests that the reported HBM synapse is a promising building block for future neuromorphic learning systems.