Background

Computation of complex, chaotic, and non-linear equations is very costly using conventional computation. Unfortunately, these types of non-linear equations frequently come up in many applications. For example, such complex, chaotic, and non-linear equations occur frequently in nature, governing the reactions of various natural phenomena (i.e., atomic particle movement, oscillation in quantum states). These natural equations are well understood, and their outputs are readable.

Reservoir computing is a method of utilizing these natural, nonlinear reactions. Reservoir computing can reduce computational costs and create systems with previously impractical levels of computational sophistication. A reservoir computer (RC) consists of three parts: i) an input layer that sends information into the physical system, ii) a dynamical system of individual units governed by non-linear reactions (the reservoir), and iii) a readout system that is trained with simple mechanisms, such as linear regression and classification.

An RC functions by sending inputs to the reservoir, allowing the natural reaction governed by complex non-linear formula to occur, and then simply observing the result. This method allows the computer to piggyback off complex, naturally occurring functions. Instead of using costly conventional methods, an RC computer uses the laws of nature to conduct the heavy lifting of the computation.

Recently, a time multiplexing scheme was introduced to build a virtual node-based RC. This development eliminates the need for a physical reservoir in the RC. The virtual nodes are created and coupled using a delayed feedback line. However, using a delayed feedback system for the virtual node formation limits the number of nonlinear systems applicable in the RC framework. Therefore, the cost and complexity of producing such an RC is high. Thus, there is substantial need for a RC that is easier and cheaper to produce.

Technology Overview

NC State researchers have developed a novel, simplified version of time-multiplexed reservoir computers (RCs) by discarding the delay and feedback lines while relying upon the physics of an oscillator to create and couple the virtual nodes. The oscillator is a forced limit-cycle oscillator (e.g., a Hopf or Lorenz oscillator). This acts as the base nonlinear dynamical system for the reservoir and can be fabricated as a circuit or other physical system.

To avoid piecewise and node dependent masks, a node independent, stochastic masking signal generated from white Gaussian noise is used. The input portion of the RC functions by modifying the external forcing and the limit cycle radius of the oscillatory circuit. A data acquisition device is used to send and receive the necessary voltages to and from the Hopf oscillator circuit. After applying a ridge regression-based training mechanism on one of the states of the Hopf oscillator reservoir, a prediction is made to follow the target signal. The Hopf oscillator RC is found to perform logical, nonlinear dynamic emulation and time series prediction tasks.

This technology could enable RCs that are cheap, robust to noise, and easy to build. Once developed, this method of RC computing is modular and can be used for any computational purposes. This will dramatically reduce computational costs across the computing industry.

Further Details

“A Hopf physical reservoir computer”, Scientific Reports (2021)- https://www.nature.com/articles/s41598-021-98982-x
“Reservoir Computing with the Lorenz System”, International Design Engineering Technical Conferences & Computers and Information in Engineering Conference (2021)-https://www.researchgate.net/publication/354933125\_RESERVOIR\_COMPUTING\_WITH\_THE\_LORENZ\_SYSTEM

Benefits

Reduces cost of RCs
Reduces noise impact in RCs

Applications

Real time audio processing
Enhanced encryption or decryption
Scientific research in computationally dense fields
Anything that requires high levels of computational power

Opportunity

NC State is looking for any partners interested in further developing and licensing this promising technology.

Reservoir Computing with Dynamic Oscillators

Seeking