Oxide-Based EDL Transistors for Neuromorphic Computing Applications

Abstract

Neuromorphic engineering, or neuromorphic computing, is a concept brought out by Carver Mead in the late eighties of last century (Mead, Analog VLSI and neural systems. Addison-Wesley Longman Publishing Co., Inc., Reading, pp 34–84 [1]). Inspired by the working patterns of human brain, neuromorphic engineering aims at building a no-bio system which has similar functions to the brain, or more specifically an ultra-low power electronic computer with powerful functions including autonomous learning and cognition (Mead, Analog VLSI and neural systems. Addison-Wesley Longman Publishing Co., Inc., Reading, pp 34–84 [1]). To realize neuromorphic engineering, scientists have proposed two methods: software-based method and hardware-based method. For software-based method, a research team in IBM once used a supercomputer named Blue Gene to simulate a neural network with the same complexity as the cat’s brain, at a rate equivalent to 1% of the real neuron response rate (Prezioso et al., In: Proceedings of the conference on high performance computing networking, storage and analysis, pp 1–12 [2]). However, this experiment cost a huge amount of computing resources, including 147,456 microprocessors, 144 TB of storage space and 1.4 MW of power. Since the software-based approach still works on the traditional sequence machines with limited parallelism, it is a great challenge for software-based methods to realize the neural morphology calculation with less resources. In contrast, the hardware-based approach can realize large-scale parallelism and high plasticity as a real neural network on the physical level. Therefore, hardware-based approach has the potential to solve this challenge.