Multimode memtransistors as optoelectronic synapses for neuromorphic computing

Abstract

Recently, a neuromorphic approach to electronics has gained attention by bringing a fundamentally different approach to existing computing architectures for pattern recognition and learning applications. Emulating complex neural behavior for a synapse through conventional Si-based devices requires many elements, increasing fabrication complexity and bringing challenges to connectivity and energy consumption. Thus, there is a need to investigate alternative material systems and device architectures for emulating neural behavior. The abrupt switching physics of most two-terminal memristors (memory + resistor) limits the number of addressable states to two/binary, limiting their plasticity and storage capacity that adversely affects the trainability of artificial neural networks. The coupling of more than one control terminal is indispensable to elicit multiple programmable conductance states and non-abrupt state transitions as weighted connections to store and update weights necessary for learning algorithms. By adopting multimodal electro-optical schemes that ‘gates’ the memconductance, one can realize multi-state optoelectronic synapses with higher order weight changes that standard two-terminal devices fail to address. This thesis explores various interfacial strategies to exploit three-terminal gate-tunable memristors (aka Memtransistors) for emulating higher-order synaptic functions.