C5-08: Spike Time Dependent Plasticity Based Spin-Neuromorphic Computing for Pattern Recognition Application. A. Lone, S. Amara and H. Fariborzi.
In this work, we proposed a spin-orbit torque-driven neuromorphic device-circuit co-design to implement the Hebbian learning algorithm. Using spike time-dependent plasticity (STDP)-based updating of synaptic weights, 8-bit, dynamic pattern recognition is illustrated. The artificial synapse we propose is based on the steady-state domain-wall motion in the MTJ free layer, driven by spin-orbit torque (SOT). The schematic of the neuromorphic circuit is presented in Fig.1. Fig 2 shows the dynamic learning and recognition capabilities of the circuit.
Fig. 1(a) Complete schematic of the neuromorphic system, showing the connection between pre-neuron and post-neuron (3T-MTJ) mediated by domain wall motion based 3T-MTJ synapse. (b) Device-circuit design on single chip for experimental realization.
Fig. 2(a) Synaptic resistance time evolution during learning phase corresponding to pattern (A) and pattern (B) respectively. Till 350 ns network completely learnt to recognize pattern (A). To show dynamic learning capability learning of pattern (B) is done. All synapses corresponding to pattern configure into lowest resistance state while background synapse attain the highest resistance state. (b) Response of post neuron after learning for input pattern and background noise. We observe that magnetization flips only for the learnt pattern.