Recently, hardware-based neural networks (HNNs) have emerged since neuromorphic systems can compute complex data efficiently. Various synaptic devices and neuron circuits suitable for architectures and learning algorithms have been researched for high...
Recently, hardware-based neural networks (HNNs) have emerged since neuromorphic systems can compute complex data efficiently. Various synaptic devices and neuron circuits suitable for architectures and learning algorithms have been researched for high performance in HNNs. Specifically, processing simultaneously both excitatory (G+) and inhibitory (G–) signals transmitted from synaptic arrays are important to process the computation efficiently and improve the performance of HNNs.
In this dissertation, synaptic and neuron devices are proposed for the neuromorphic system with high density and low power consumption. A positive-feedback (PF) device simultaneously processing excitatory and inhibitory signals is used as the neuron device to replace conventional neuron circuits. Owing to the steep switching characteristics of the PF operation, the PF neuron device can reduce the energy consumption during processing integration function of neurons. The PF neuron device is an efficient structure that merges a gated thyristor and a single MOSFET. By accumulating electrons in an n floating body of the PF neuron device, the integrate-and-fire operation with steep subthreshold swing (SS < 1 mV/dec) is experimentally implemented. The electrons accumulated in the n floating body are discharged by applying inhibitory signals to the merged FET. Moreover, the threshold voltage (Vth) of the proposed PF neuron with a non-volatile memory function is controlled by program and ease states in a charge storage layer. The PF neuron circuit that consumes low energy per a spike (~ 0.62 pJ/spike) consists of one PF device and only five MOSFETs for the integrate-and-fire function and reset operation. The dual-gate FET with independent two gates (G1 and G2) is proposed as the synaptic device. Here, G1 turns on and off the synaptic device, and G2 with the charge storage layer controls the conductance of the dual-gate FET for synaptic weights. The range of conductance change of the dual-gate FET is very wide (100 pA ~ 1 μA). In the NOR type array based on the dual-gate FETs, program and erase operations can be implemented with the Fowler-Nordheim (FN) tunneling mechanism, resulting in low power consumption during the synaptic weight update. The sum of current (3.63 μA) of eight individual dual-gate FETs is almost the same (~ 0.87 %) as the Itotal (3.6 μA) of eight dual-gate FETs in the NOR type synapse array. The variations (σ/μ) of the quantized synaptic currents in eight synaptic devices are obtained as 0.023, 0.011, 0.015, and 0.032 for four different synaptic weight states. The PF neuron circuit and synapse array based on the dual-gate FETs provide viable solutions for high-density and low-energy neuromorphic systems.