A novel Reservoir Architecture for Periodic Time Series Prediction

Zhongju Yuan,Geraint Wiggins,Dick Botteldooren

2024-05-16

Abstract:This paper introduces a novel approach to predicting periodic time series using reservoir computing. The model is tailored to deliver precise forecasts of rhythms, a crucial aspect for tasks such as generating musical rhythm. Leveraging reservoir computing, our proposed method is ultimately oriented towards predicting human perception of rhythm. Our network accurately predicts rhythmic signals within the human frequency perception range. The model architecture incorporates primary and intermediate neurons tasked with capturing and transmitting rhythmic information. Two parameter matrices, denoted as c and k, regulate the reservoir's overall dynamics. We propose a loss function to adapt c post-training and introduce a dynamic selection (DS) mechanism that adjusts $k$ to focus on areas with outstanding contributions. Experimental results on a diverse test set showcase accurate predictions, further improved through real-time tuning of the reservoir via c and k. Comparative assessments highlight its superior performance compared to conventional models.

Neural and Evolutionary Computing,Artificial Intelligence,Machine Learning,Audio and Speech Processing

What problem does this paper attempt to address?

The problem this paper attempts to address is periodic time series prediction, particularly its application in music rhythm prediction. Specifically, the authors propose a new architecture based on reservoir computing, aiming to accurately predict rhythm signals within the range of human perception. This issue holds significant importance in various fields such as industrial processes and music generation. Although traditional machine learning and statistical models can perform periodic time series prediction, they often lack sufficient generalization ability and may require retraining for data with different frequencies. Additionally, existing deep learning models, while capable of completing this task, require substantial computational resources and are limited in real-time hardware applications due to an excessive number of parameters.

To overcome these issues, the authors chose the reservoir computing method, which has significant advantages in training efficiency because reservoir weights typically do not need to be trained. By designing a new reservoir structure, the authors hope that the model can simulate human perception and synchronization abilities for rhythm and beats while maintaining a low number of parameters, making it more suitable for implementation on practical hardware. Furthermore, the authors introduce a new loss function and a dynamic selection (DS) mechanism to further improve the model's prediction accuracy and synchronization performance.