Autoassociative Memory and Pattern Recognition in Micromechanical Oscillator Network

Ankit Kumar,Pritiraj Mohanty

doi:10.1038/s41598-017-00442-y

Ankit Kumar, Pritiraj Mohanty

Open Access

https://doi.org/10.1038/s41598-017-00442-y

Copy DOI

Abstract

Towards practical realization of brain-inspired computing in a scalable physical system, we investigate a network of coupled micromechanical oscillators. We numerically simulate this array of all-to-all coupled nonlinear oscillators in the presence of stochasticity and demonstrate its ability to synchronize and store information in the relative phase differences at synchronization. Sensitivity of behavior to coupling strength, frequency distribution, nonlinearity strength, and noise amplitude is investigated. Our results demonstrate that neurocomputing in a physically realistic network of micromechanical oscillators with silicon-based fabrication process can be robust against noise sources and fabrication process variations. This opens up tantalizing prospects for hardware realization of a low-power brain-inspired computing architecture that captures complexity on a scalable manufacturing platform.

Highlights

The dynamics of a system of coupled oscillators can exhibit attractive limit cycles that represent synchronized states
As the required fabrication methods are based on standard lithography and processing techniques currently in use at semiconductor foundries, any architecture based on an array of MEMS oscillators would lend itself to highly scalable manufacturing
We present detailed simulation results based on a realistic physical model that captures the dynamics of these resonators

Summary

The frequencies

The stability of the set of fixed points aio of the complex amplitudes implies stability of limit cycles for the original displacement x This existence of such a set can be proven without explicitly solving for the dynamical trajectories through the construction of a Lyapunov function, which can be considered to be a generalized energy function for the system. Such network segmentation in the presence of incommensurate frequencies can greatly hinder phase synchronization, since phase locking requires frequency entrainment, yet such entrainment in general is impossible without at least indirect coupling between every oscillator This somewhat limited storage capacity is a fundamental limitation of our reliance on dynamical fixed points

Absence of Noise

Addition of Noise

Micromechanical Self Oscillators

Conclusion

Additional Information