Artificial neurosynaptic device based on amorphous oxides for artificial neural network constructing

Abstract

The von Neumann architecture places restrictions on how much energy conventional computers can use for artificial intelligence training. We created a three-terminal artificial neural synapse that is electronically controlled using amorphous InAlZnO, which was inspired by biological synapses. With electrical pulses of −5 V and 5 V, respectively, the device can produce excitatory postsynaptic current (EPSC) and inhibitory postsynaptic current (IPSC), demonstrating good biological synaptic characteristics. A single pulse can use as little as 456.69 pJ of energy. Meanwhile, the synaptic device will transition from the short-term plasticity (STP)/short-term depression (STD) state to the long-term plasticity (LTP)/long-term depression (LTD) state as the number of pulses grows, more closely mimicking the features of brain learning and memory. Furthermore, conductance stability of the synaptic device is maintained after several LTP–LTD cycles, and we built a three-layer artificial neural network (ANN) on the basis of this attribute. Its handwritten digit recognition accuracy reaches 92.26% after training it with the MINST dataset. The amorphous oxide artificial neural synapse device developed in this work will be crucial in addressing the issue of artificial intelligence training's energy consumption as well as the development of artificial neural networks in the future.