Solving optimization tasks power-efficiently exploiting VO₂'s phase-change properties with Oscillating Neural Networks

Abstract

Recent progress in the development of phase-change materials is enabling their use for novel approaches in spike-based learning circuits and brain-inspired computing architectures [1], [2]. The race to leverage efficient low-power neural processing systems whose fundamental operations are akin to those of animal brains drives the community to explore new 'neuromorphic' materials and devices [3]. Amongst the candidates suitable to deliver such types of devices, vanadium-dioxide (VO2) offers promising features [4]. Its polycrystalline morphology, once integrated on a Si platform, has the ability to oscillate from a high to a low resistive statenear room temperature (68°C) [5]. These natural oscillations triggered by biasing VO2 at low voltage provide means to build a network of electronic oscillators with tremendous potential for neural network architectures, AI applications, and optimization capabilities [1], [3], [6], [7]. In particular, oscillation-based computing serves best to solve constraint-satisfaction problems [8], [9]. Any optimization problem, simple in appearance, typically requires heavy computing resources calling for long processing times and even larger energy consumption [2]. In this work, we show the power and the number of cycles needed to reach a solution can be dramatically reduced to solve NP-hard problems with our befitted VO2-based 3\times 3 oscillating neural networks (ONNs).