ScholarWorks@동국대학교

초록

Synaptic emulation using memristive devices has advanced neuromorphic computing by enabling energy-efficient and scalable architectures. Here, we report a non-volatile TiN/Al/TiN/Ti/TiOx/Al2O3/Pt resistive random-access memory (RRAM) device featuring an oxygen-deficient TiOx switching layer. The device exhibits reliable long-term memory characteristics with stable multi-level current modulation. Neuromorphic functionalities such as pattern learning and classification using the EMNIST dataset, as well as 4-bit edge computing, are successfully demonstrated, with the classification achieving an accuracy of 91.18%. While prior studies predominantly focused on excitatory synaptic behaviors, this work introduces a hardware-level approach to emulate synaptic forgetting, an essential but underexplored aspect of biological memory processing. To implement forgetting, we propose three experimental methodologies: (1) inhibitory postsynaptic current (IPSC) for synaptic suppression, (2) reversed Pavlovian conditioning to emulate de-learning, and (3) activity-dependent synaptic selection (ADSS) mimicking biologically realistic synaptic pruning. These strategies enable selective synaptic weakening based on firing strength and frequency, closely resembling natural forgetting mechanisms. Our findings establish a new paradigm in neuromorphic hardware that balances learning and forgetting using non-volatile devices. This direction not only enhances biological plausibility but also broadens the functional capabilities of memristive systems for adaptive and efficient edge AI applications.

키워드