A Hopfield neural network in magnetic films with natural learning

In order to develop its ultimate potential, artificial intelligence must shift from the software-based towards hardware-based neuromorphic algorithms. For the artificial neural network, it requires materials and devices that are capable of emulating neuron-like activation and synapses with reprogrammable weights. To the best of our knowledge, all existing approaches in implementing synapses require external computation to train the network. It would be highly desired that the hardware can naturally learn by itself with less or even without external intervention. 

We distinguish two stages for natural learning based on physical systems: i) the physical systems CAN be adjusted via external stimulus; ii) the physical systems KNOW HOW to adjust. Most physical systems, including the memristor, can realize stage i), i.e. the stored weights in memristor can be modified via external stimulus. However, most physical systems don’t know how to adjust, so they need algorithms (such as the back-propagation algorithm) carried out by external computers to calculate the amount of the adjustment.

Screen Shot 2021-11-19 at 2.35.08 PM

In this paper, we propose a hardware scheme that is capable in realizing stage two natural learning, i.e. the hardware itself knows how to adjust according to inputs. Here we use conducting magnetic films with inhomogeneous textures as a synaptic network. The weights are stored as a conductance matrix defined by distributed electric contacts. Beacuse of the plasticity of the magnetic texture (see figure on the right), electric currents can train the network by modulating the texture via the current-induced spin transfer torque. Different from conventional implementations, the malleable magnetic textures autonomously train the network according to the Hebbian learning principle. The synaptic “plasticity” and “competence” required by Hebb’s rule is built in intrinsically in the physical process. As a proof of principle, we numerically simulate the on-chip training and inferring of a 4-node Hopfield network (see figure below).

Screen Shot 2021-11-19 at 2.16.37 PM

This work was mainly conducted by Dr. Weichao Yu, who is now a faculty member of the Institute for Nanoelectronic devices and Quantum computing, Fudan University. This work was carried out in collaboration with Prof. Gerrit E. W. Bauer from Tohoku University, Japan.

[1]. Hopfield neural network in magnetic textures with intrinsic Hebbian learning
Weichao Yu, Jiang Xiao*, and Gerrit E. W. Bauer
Phys. Rev. B 104, L180405 (2021)

© Jiang Xiao 2014