Sub-harmonic Injection Locking Memristor-based Oscillator Arrays Used for Pattern Recognition

Abstract

Oscillatory neural network (ONN) is a promising computing architecture that can realize pattern recognition or other intellectual applications in real time. Emerging memristor-based oscillator provides a good choice as building block and corresponding phase behavioral model can accelerate simulation for about 40 times comparing to transistor level simulation. However, phase unlocking due to the frequency detuning after recognition is prominent in memristor-based traditional ONN architecture. With such problem, phase deviation cannot be locked and the correct recognized image fails to be continuously presented after synchronization. In this paper, a novel sub-harmonic injection locking (SHIL) memristor-based ONN is proposed to handle this problem. Energy function is used to give a deep insight into our method. Simulation both in MATLAB and Cadence shows a consistent frequency, i.e., constant phase differences. The results of error standard deviation of recognized patterns reduce 80, 36 and 192 times respectively in three representative cases.