Silicon-Based Synaptic Transistor for Neuromorphic Computing Systems

Abstract

Current computing systems based on the von Neumann architecture suffer from the fact that serious leakage current issues have risen up in nanoscale devices. Neuromorphic computing system has been believed to solve fundamental challenges in current computing system by mimicking a biological nervous system, especially in terms of parallel signal processing. Synaptic devices are considered as the one of the most important parts of neuromorphic systems because a biological synapse is thought to control signal transmissions and memory effects in our nervous system. However, the memristor, one of the strongest candidates for an artificial synapse, requires additional switching parts in order to transfer and receive signals using the same electrode, leading to extra overheads that may compromise the advantages of massive parallelism inherent in neuromorphic systems. In this dissertation, a silicon-based synaptic transistor with asymmetric dual-gate structure is investigated. The structural feature enables the synaptic transistors to interact with both pre- and post-synaptic neuron circuits directly. A TCAD device simulator and a circuit simulator are used to verify its synaptic learning properties and study its mechanism fundamentally. After verifying all the fabrication flow using a process simulator, the synaptic transistors are fabricated through process techniques including two-step CMP processes. The electrical and synaptic characteristics of the fabricated devices are measured with a semiconductor parameter analyzer, and a device model is created based on the measured data. Furthermore, spiking neural network composed of them is verified systematically using the device model. From the simulation study and electrical measurement, synaptic learning rules are observed in the synaptic transistors including the transition from short-term to long-term memory and spike-timing dependent plasticity. In addition, the spiking neural network composed of the synaptic transistors boasted its ability of pattern recognition using MNIST data set. The total recognition accuracy of the hardware-based neural network system having 784 input nodes and 10 output nodes is improved to nearly 70% by adding inhibitory synapses. These results indicate that the synaptic transistor studied in this dissertation can be used as a synaptic device in neuromorphic systems thanks to its direct connectability with neuron circuits and synaptic learning properties.