In situ unsupervised learning using stochastic switching in magneto-electric magnetic tunnel junctions

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) offer bio-plausible, power-efficient alternatives to deep learning.
• Current hardware implementations struggle to match the brain's power efficiency.
• Novel materials are needed to effectively mimic SNN components like neurons and synapses.

Purpose of the Study:

• To introduce a magnetoelectric-magnetic tunnel junction (ME-MTJ) device as a functional synapse for SNNs.
• To demonstrate in situ unsupervised learning using ME-MTJ synapses in a crossbar array.
• To achieve energy-efficient SNN implementation for AI tasks.

Main Methods:

• Utilized the capacitive nature of ME-MTJ write-ports for probabilistic switching via voltage pulses.
• Exploited sigmoidal switching characteristics to implement a stochastic spike-timing-dependent plasticity (STDP) rule.
• Simulated a two-layered SNN using ME-MTJ synapses for image classification.

Main Results:

• Successfully implemented a stochastic STDP rule in ME-MTJ synapses.
• Demonstrated the feasibility of transistor-less crossbar arrays for SNNs.
• Achieved image classification on a handwritten digit dataset using the developed SNN.

Conclusions:

• ME-MTJ devices serve as efficient synapses for SNNs.
• The decoupled read-write path and capacitive write-port enable transistor-less, energy-efficient crossbar arrays.
• This approach facilitates in situ learning in SNNs, advancing brain-inspired computing.