A New Spin on Machine Learning

Spintronic computing devices represent a cutting-edge approach to information processing that harnesses the intrinsic spin property of electrons in addition to their charge. Unlike traditional electronic devices that rely solely on electron charge to encode information, spintronics utilizes the electron spin, which is a quantum mechanical property related to the intrinsic angular momentum of the electron. This spin can be either "up" or "down," serving as the basis for encoding information.

At the heart of spintronic devices are spintronic elements, such as magnetic tunnel junctions, spin valves, and spin-transfer torque devices. These components exploit the manipulation of electron spins through magnetic fields and allow for the generation, transport, and detection of spin-polarized currents. Spintronics not only enhances the energy efficiency of information processing but also holds promise for novel functionalities and devices that surpass the limitations of conventional electronics.

One intriguing concept within spintronics is the use of probabilistic bits, or p-bits. Traditional bits in classical computing are binary, representing either a 0 or a 1. P-bits, on the other hand, leverage the probabilistic nature of electron spin states, allowing them to exist in a superposition of both states simultaneously. This inherent probabilistic behavior makes p-bits particularly suitable for addressing complex computational problems, such as optimization and cryptography.

A team at Tohoku University and the University of California, Santa Barbara realized that as Moore's Law continues to slow down, there will be an increasing need for specialized hardware. The present boom in machine learning, in particular, is making this very apparent. Seizing on this opportunity, they have developed a proof of concept probabilistic spintronic computing system that is well-suited for solving many computationally hard tasks in machine learning.

The researchers’ primary innovation was in the use of a spintronic device called a stochastic magnetic tunnel junction (sMTJ) in concert with a field programmable gate array to create a room-temperature probabilistic computer.

In the past, sMTJ-based probabilistic computers were only demonstrated as being capable of implementing recurrent neural networks. While these models are quite useful for a variety of applications, feedforward neural networks power the majority of modern artificial intelligence applications, excluding probabilistic computing platforms. In this recent work, the research team demonstrated that their platform could run feedforward neural networks.

To accomplish this goal, the new spintronic computing system leverages a new type of sMTJ. These sMTJs allowed for the development of p-bits that can change their state about a million times per second, which is orders of magnitude faster than any similar p-bit technology. Second, they devised a method to arrange and update these p-bits in a specific order, allowing them to function together like building blocks. As a result, they demonstrated how this technology can execute a fundamental operation of a Bayesian network, a sort of computational model frequently utilized in machine learning. In essence, they have shown that their new technology can function as part of a system that mimics how our brains process information in certain types of calculations.

Current experiments have been small-scale, but the team hopes to scale up their system in the future, perhaps with the help of advancements in CMOS-compatible Magnetic RAM technology.