Synaptic Plasticity in Memristive Artificial Synapses and Their Robustness Against Noisy Inputs

Area of Science:

• Neuromorphic Computing
• Materials Science
• Artificial Intelligence Hardware

Background:

• Emerging neuromorphic computing requires brain-inspired devices mimicking biological synapses for efficient, cognitive data processing.
• Memristive devices are promising for artificial synapses due to their multilevel and dynamical plastic behaviors.
• BiFeO3 (BFO)-based memristive devices offer ultrastable analog characteristics for synaptic emulation.

Purpose of the Study:

• To experimentally demonstrate the functionality of BFO artificial synapses in emulating long-term synaptic plasticity.
• To investigate the impact of electrical stimuli on spike-timing-dependent plasticity (STDP) and explore frequency-dependent plasticity (FDP).
• To analyze the robustness of BFO artificial synapses against intrinsic and extrinsic neuronal noise.

Main Methods:

• Utilized ultrastable analog BiFeO3 (BFO)-based memristive devices.
• Experimentally demonstrated spike-timing-dependent plasticity (STDP), cycle number-dependent plasticity (CNDP), and spiking rate-dependent plasticity (SRDP).
• Analyzed the influence of pulse width and amplitude on STDP and investigated generalized frequency-dependent plasticity (FDP) by modulating pulse width and interval ratios.

Main Results:

• BFO artificial synapses exhibit diverse long-term plastic functions (STDP, CNDP, SRDP) and tunable learning windows.
• Frequency-dependent plasticity (FDP) reveals that ratio modulation between pulse width and interval can induce both potentiation and depression at the same frequency.
• BFO artificial synapses demonstrate robustness against intrinsic noise, with extrinsic noise impact on STDP analyzed.

Conclusions:

• Ultrastable BFO memristive devices effectively emulate artificial synapses with versatile plasticity.
• The tunable nature of BFO synapses, including FDP, is crucial for advanced neuromorphic applications.
• BFO artificial synapses show significant potential for building robust and efficient brain-inspired computing systems.