A ferrofluid based neural network: design of an analogue associative memory

TL;DR

This paper proposes a novel analogue associative memory using ferrofluid-based magnetic particles, trained through energy minimization, capable of high-accuracy pattern storage and recall demonstrated via simulations.

Contribution

It introduces a ferrofluid-based neural network design with new training algorithms and demonstrates its effectiveness through Monte Carlo simulations.

Findings

Capable of storing and recalling two pattern sets with near 100% accuracy.

Uses energy minimization during training to form stable memory states.

Employs spin-valve sensors for magnetic readout of memory states.

Abstract

We analyse an associative memory based on a ferrofluid, consisting of a system of magnetic nano-particles suspended in a carrier fluid of variable viscosity subject to patterns of magnetic fields from an array of input and output magnetic pads. The association relies on forming patterns in the ferrofluid during a trainingdphase, in which the magnetic dipoles are free to move and rotate to minimize the total energy of the system. Once equilibrated in energy for a given input-output magnetic field pattern-pair the particles are fully or partially immobilized by cooling the carrier liquid. Thus produced particle distributions control the memory states, which are read out magnetically using spin-valve sensors incorporated in the output pads. The actual memory consists of spin distributions that is dynamic in nature, realized only in response to the input patterns that the system has been trained for. Two training algorithms for storing multiple patterns are investigated. Using Monte Carlo simulations of the physical system we demonstrate that the device is capable of storing and recalling two sets of images, each with an accuracy approaching 100%.