Robust autoassociative memory with coupled networks of Kuramoto-type oscillators

Abstract

Uncertain recognition success, unfavorable scaling of connection complexity, or dependence on complex external input impair the usefulness of current oscillatory neural networks for pattern recognition or restrict technical realizations to small networks. We propose a network architecture of coupled oscillators for pattern recognition which shows none of the mentioned flaws. Furthermore we illustrate the recognition process with simulation results and analyze the dynamics analytically: Possible output patterns are isolated attractors of the system. Additionally, simple criteria for recognition success are derived from a lower bound on the basins of attraction.

Noise Immunity of Oscillatory Computing Devices

“…OBCs have several other, possibly game-changing benefits such as the possibility to use frequency-domain multiplexing for neuron-to-neuron communication 24,25 or device-level advantages, such as realization with very compact, low-power devices . These were not discussed in this paper and are in addition to the noise-related strengths.…”

What are higher-order networks?

“…These and other higher-order terms in the Kuramoto model can have a range of dynamical effects and are directly relevant to understand the dynamics of oscillations since they arise in a number of physical systems. Examples include oscillatory dynamics in electronic circuits [129], the dynamics of coupled nanomechanical oscillators [181], and optical devices [214]. Apart from an impact on synchronization [159,174,244], higher-order interactions can lead to multistability [258], heteroclinic cycles [28,29,33], and chaotic dynamics [30].…”

Perspectives of using spin waves for computing and signal processing

“…There are a few recent ideas for oscillatory neurons that are a lot more (nano)technology friendly. The work of [92] uses the a frequency-domain multiplexing scheme to drastically reduce the number of inteconnections. Along a similar line of thought, the work of [105] uses external oscillatory signals to control partial synchronization of oscillators and effectively control their interconnection strengths this way.…”

“…Microelectronic technologies cannot even come close to this number. One way to create a highly interconnected network is to use frequencydivision multiplexing (FDM) in artificial neural networks and use a single, high-bandwidth physical link to create a large number of channels between processing units (neurons) [88], [89] [90] [91] [75] [92] [79] [93]. The possibility of using FDM provides another hand-waving argument in favor of oscillatory signal representation.…”