3T-1C-1R Synaptic Design for Online Learning in Spiking Neural Networks

Recurrent Neural Networks (RNNs) are widely used in diverse applications such as Natural Language Processing (NLP), speech synthesis, and autonomous systems due to their ability to efficiently process complex sequential data. Their recurrent architecture, akin to lateral recurrent connections in the neocortex, enhances expressiveness but poses challenges in training, particularly under constrained computational resources. A key issue in RNNs is the credit assignment problem, which determines synaptic weight updates and is traditionally addressed using backpropagation through time (BPTT) [1]. However, BPTT demands significant memory and computational power, limiting its feasibility for neuromorphic hardware. In Spiking Neural Networks (SNNs), alternative biologically inspired approaches, such as eligibility traces, dendritic processing, and neuromodulatory signals, provide more efficient learning mechanisms [2]. This work explores novel synaptic designs that enable eligibility trace-based online learning in SNNs.