An Oscillatory Neural Network Based Local Processing Unit for Pattern Recognition Applications

Abstract

Prolific growth of sensors and sensor technology has resulted various applications in sensing, monitoring, assessment and control operations. Owing to the large number of sensing units the the aggregate data volume creates a burden to the central data processing unit. This paper demonstrates an analog computational platform using weakly coupled oscillator neural network for pattern recognition applications. The oscillator neural network (ONN) has been studied over the last couple of decades for it’s increasing computational efficiency. The coupled ONN can realize the classification and pattern recognition functionalities based on its synchronization phenomenon. The convergence time and frequency of synchronization are considered as the indicator of recognition. For hierarchical sensing, the synchronization is detected in the first layer, and then the classification is accomplished in the second layer. In this work, a Kuramoto model based frequency synchronization approach is utilized, and simulation results indicate less than 160 ms convergence time and close frequency match for a simplified pattern recognition application. An array of 10 sensors is considered to affect the coupling weights of the oscillating nodes, and demonstrate network level computation. Based on MATLAB simulations, the proposed ONN architecture can successfully detect the close-in-match pattern through synchronization, and differentiate the far-out-match pattern through loss of synchronization in the oscillating nodes.

Coupled VO2 Oscillators Circuit as Analog First Layer Filter in Convolutional Neural Networks

“…Another popular configuration encodes the distance between the input pattern and the feature to be recognized in a frequency shift between oscillators, which are connected by a fixed coupling. The distance between the two patterns is then calculated on the time the oscillators need to converge to the same frequency ( Cotter et al, 2014 ; Nikonov et al, 2015 ; Zhang et al, 2019 ). The implementation of these concepts does not use the associative memory capabilities of the ONNs to store multiple patterns.…”

Current Modulation of Nanoconstriction Spin-Hall Nano-Oscillators

“…[42] • Oscillatory arrays with multilevel neurons for pattern recognition in the frequency domain. [43,44] • Auto-associative memory and pattern recognition with an oscillator network; this involves looking at the strength of synchronization defined by the output power, in which tuning the phases (the coupling in the Kuramoto model) will change the synchronization strength. [45][46][47] • Oscillatory network based on a Kuramoto phase oscillator for image segmentation.…”

“…In two-dimensional SHNO arrays, an individual SHNO operating at a gigahertz frequency is coupled to neighboring SHNOs, imitating the oscillatory behavior of neurons connected to their neighbors via synapses. There are proposals to use coupled oscillators network as ultrafast and efficient non-Von Neumann computing paradigms for a range of applications [42][43][44][45][46][47][48][49][50][51]. Implementing such paradigms will transform spintronics in a technology capable of providing all the pieces of the IoT puzzle, from fast memory to ultra-compact communication systems and processing units.…”

Coupled Moebius Maps as a Tool to Model Kuramoto Phase Synchronization

“…For instance, models of coupled phase oscillators have been successfully applied to functional connectivity of the human brain [3,4], neuronal oscillatory behaviors [5,6], and neural encoding [7][8][9]. Increasingly, they also serve as a computational tool in machine learning and artificial intelligence based on oscillatory neural networks [10][11][12][13], which opens up a new perspective for hardware implementations [14].…”