Neuromorphic Hebbian learning with magnetic tunnel junction synapses.

Area of Science:

• Materials Science
• Computer Science
• Neuroscience

Background:

• Neuromorphic computing seeks to emulate biological neural networks for energy-efficient AI.
• Conventional neuromorphic systems face challenges with analog memory states, including stochasticity and limited endurance.
• Efficiently storing synaptic weights is crucial for in-memory computation in neuromorphic architectures.

Purpose of the Study:

• To propose and demonstrate neuromorphic networks utilizing binary resistance states of magnetic tunnel junctions (MTJs) for high-accuracy inference.
• To leverage the analog stochastic switching of spin-transfer torque (STT) in MTJs for unsupervised Hebbian learning.
• To explore a hardware-aware design for STT-MTJ neuromorphic learning networks enabling autonomous AI.

Main Methods:

• Experimental implementation of neuromorphic networks with MTJ synapses for inference and learning.
• Utilizing binary resistance states of MTJs for synaptic weight storage.
• Leveraging stochastic spin-transfer torque switching in MTJs for unsupervised Hebbian learning.

Main Results:

• Demonstrated high-accuracy inference in experimental neuromorphic networks using MTJ synapses.
• Successfully implemented spike-timing-dependent plasticity learning with MTJ synapses.
• Simulations showed competitive MNIST recognition accuracy for unsupervised Hebbian learning with STT-MTJ synapses.

Conclusions:

• STT-MTJ based neuromorphic networks offer a promising approach for efficient and autonomous AI.
• Binary resistance states of MTJs ensure high-precision inference, while stochastic switching enables effective learning.
• Hardware-aware design of these networks facilitates the development of next-generation AI hardware.