Influence of Thermal Fluctuations on Self-Organization in Coupled Oscillatory Neural Networks

The rise of AI and increasing computational demands are leading to unsustainable energy consumption in computing systems. To address the urgent need for energy-efficient computing, novel physics-inspired computing models are being explored. Computing with coupled oscillators, or oscillatory neural networks (ONNs), represents a novel, physics-inspired computing paradigm. Originating from Hopfield neural networks, ONNs harness the complex dynamics of coupled oscillators to perform computations through phase synchronization while minimizing their energy. ONN computing is based on the synchronization phenomena of coupled oscillators, which can be described by the Kuramoto model. This model captures the evolution of oscillator phase dynamics over time as a function of coupling strengths.

Typically, such systems are studied for varying coupling strengths and oscillator models. Here, we report that thermal fluctuations also play a crucial role in modulating coupling strengths, facilitating the self-organization of the network and the emergence of richer oscillator phase states. Interestingly, thermal fluctuations occur at a slower timescale than the oscillator switching frequencies, leading to an agglomeration of dynamics across different timescales. We observe that thermal fluctuations introduce nonlinear dynamics, significantly impacting the performance of ONNs in associative memory tasks such as pattern recognition.