A Study of Changes in Prediction Performance Influenced by Attractor State in Oscillator Reservoir Computing

AI technologies have proven their capabilities in solving complex problems, but they often require significant computational resources and power. This study proposes a reservoir computing (RC) model utilizing Bonhoeffer-van der Pol (BVP) oscillators, leveraging their chaotic dynamics to generate complex, high-dimensional internal states. The paper investigates the impact of the attractor states of these oscillators on the prediction performance of a speech recognition task. Results indicated that the asymmetric attractors led to lower Word Error Rates (WER) compared to periodic ones, suggesting that the diversity of circuit responses is beneficial for RC performance.