Neuromorphic Hebbian learning with magnetic tunnel junction synapses

Abstract: Neuromorphic computing aims to mimic both the function and structure of biological neural networks to provide artificial intelligence with extreme efficiency. Conventional approaches store synaptic weights in non-volatile memory devices with analog resistance states, permitting in-memory computation of neural network operations while avoiding the costs associated with transferring synaptic weights from a memory array. However, the use of analog resistance states for storing weights in neuromorphic systems is impeded by stochastic writing, weights drifting over time through stochastic processes, and limited endurance that reduces the precision of synapse weights. Here we propose and experimentally demonstrate neuromorphic networks that provide high-accuracy inference thanks to the binary resistance states of magnetic tunnel junctions (MTJs), while leveraging the analog nature of their stochastic spin-transfer torque (STT) switching for unsupervised Hebbian learning. We performed the first experimental demonstration of a neuromorphic network directly implemented with MTJ synapses, for both inference and spike-timing-dependent plasticity learning. We also demonstrated through simulation that the proposed system for unsupervised Hebbian learning with stochastic STT-MTJ synapses can achieve competitive accuracies for MNIST handwritten digit recognition. By appropriately applying neuromorphic principles through hardware-aware design, the proposed STT-MTJ neuromorphic learning networks provide a pathway toward artificial intelligence hardware that learns autonomously with extreme efficiency.

• The paper introduces a novel neuromorphic architecture using MTJ synapses that combine stable binary resistance states with stochastic switching for Hebbian learning.
• Experimental results show accurate image recognition via a 4x2 MTJ network that performs reliable vector-matrix multiplication with binary inputs.
• Simulations confirm scalability and high inference accuracy on tasks like MNIST digit recognition, highlighting the approach's potential for advanced AI hardware.

Neuromorphic Hebbian Learning with Magnetic Tunnel Junction Synapses

The paper "Neuromorphic Hebbian learning with magnetic tunnel junction synapses" focuses on leveraging spin-transfer torque (STT) magnetic tunnel junction (MTJ) synapses for energy-efficient neuromorphic computing and unsupervised Hebbian learning. This approach aims to mimic biological neural networks both structurally and functionally while overcoming challenges associated with conventional analog memory devices such as stochastic writing and weight drift.

Introduction to MTJ-Based Neuromorphic Networks

Magnetic tunnel junctions (MTJs) are proposed as binary resistance devices for neuromorphic computing due to their stable resistance states. They facilitate more accurate neural network inferences compared to analog memory devices suffering from stochasticity and endurance issues. While MTJs naturally offer stable binary states, their stochastic switching behavior through spin-transfer torque (STT) provides the analog characteristics necessary for unsupervised Hebbian learning. This offers potential for new applications in autonomous AI systems that require localized learning without labeled data.

Experimental Demonstration of MTJ Binary Neuromorphic Network

The experimental demonstration of MTJ-based neuromorphic networks provides insight into practical applications of these devices. A 4x2 MTJ neuromorphic network was implemented to showcase image recognition capabilities using 2x2 pixel images. By encoding synaptic weights in MTJ conductance states, the network performs vector-matrix multiplication (VMM) operations with binary input voltages.

Figure 1: MTJs for binary neuromorphic computing. a, Vector-matrix multiplication (VMM) with a resistive memory crossbar array. The matrix values are encoded in the device conductances, G. The input vector is converted to an array of voltages fed into the column lines, $\vec{v}_{in}$.

This experimental setup verified the MTJ network's ability to accurately classify images based on Hamming distances between input and target images, demonstrating the network's fidelity to expected VMM results. The demonstration highlights the networks' potential utility in AI systems with reduced variation expected in future implementations.

Unsupervised Hebbian Learning with STT-MTJ Stochastic Switching

Unsupervised Hebbian learning, based on the classical Hebbian postulate, emphasizes learning through local activity—simple processes where neurons that activate simultaneously strengthen their connections. This paper proposes a neuromorphic system using STT-MTJ synapses that combines stable binary states with stochastic properties for efficient unsupervised learning.

Figure 2: Unsupervised learning with stochastic STT-MTJ switching. a-b, Backpropagation vs. localized Hebbian learning.

Unlike complex backpropagation methods, this approach employs Hebbian learning rules facilitated by STDP, shown to be more computationally efficient and easily implemented in hardware. The stochastic switching probability characteristic of STT-MTJs is leveraged for analog learning processes, allowing the network to update synaptic weights based on relative neuron firing activity.

Experimental Demonstration of Unsupervised Hebbian Learning

A 4x2 STT-MTJ network was utilized to demonstrate unsupervised Hebbian learning in practice. The experiment involved a manual control switch platform that simulated neuron firing and updated synaptic states through STDP pulses. Successful learning was demonstrated when neurons specialized in recognizing specific input patterns after training with unsupervised learning.

Figure 3: Experimental demonstration of unsupervised Hebbian learning with stochastic STT-MTJ switching. a, Schematic of the 4x2 neuromorphic learning network.

Over several input presentations, neurons adapted to recognize distinct images through stochastic MTJ state changes prompted by learning pulses. This experiment verified the feasibility of applying stochastic STT-MTJ switching for robust unsupervised learning tasks in neuromorphic networks.

Scaling Hebbian Learning to Large Networks

Simulations validate the scalability of STT-MTJ learning systems to larger networks, particularly through MNIST digit recognition tasks. They demonstrate the architecture’s ability to achieve high inference accuracy with increased neuron and synapse counts, despite the challenges associated with memristors and PCM devices.

Figure 4: Behavioral simulation of unsupervised Hebbian learning of the MNIST handwritten digit dataset with stochastic STT-MTJ switching.

These simulated results underscore the advantages of binary STT-MTJs over multilevel devices by achieving competitive accuracies without compromising practical viability or assuming impractical device behavior.

Conclusions

This paper presents a promising framework for neuromorphic computing using magnetic tunnel junction synapses, offering practical solutions for high-efficiency AI hardware. The demonstrated capabilities in both inference and learning suggest a pathway toward scalable, autonomous systems utilizing unsupervised Hebbian learning principles. Such systems promise to expand AI applications significantly, especially in areas demanding efficient processing and adaptation to novel data. Through leveraging the distinct properties of STT in MTJs, this approach offers a robust and efficient architecture for future neuromorphic intelligence devices.