Bio-inspired learning algorithm for time series using Loewner equation

TL;DR

This paper introduces bio-inspired learning methods for time series analysis using the Loewner equation, combining statistical mechanics and neural dynamics, with numerical tests on neuronal models.

Contribution

It proposes two novel learning algorithms based on Loewner theory for time series, linking biological information processing and nonlinear dynamics.

Findings

Gaussian process regression applied to Loewner driving forces shows effective time series modeling.

Fluctuation dissipation relation derived from Loewner theory measures system sensitivity.

Numerical tests on neuronal models demonstrate the methods' applicability.

Abstract

Though the relationship between the theoretical statistical physics and machine learning techniques has been a well-discussed topic, the studies on the mechanism of learning inspired by the biological system are still developing. In this study, we investigate the application methods of Loewner equation to the learning algorithm particularly focusing on its statistical-mechanical aspects. We suggest two simple methods of learning of one dimensional time series based on the unique encoding property of the discrete Loewner evolution. The first one is the Gaussian process regression using the normality of the distribution of Loewner driving force corresponding to the curve composed from the time series. The second one is the fluctuation dissipation relation for the time series, which is derived from the Loewner theory, measuring the sensitivity of the nonlinear dynamics under the small perturbation. These methods were numerically tested dealing with the neuronal dynamics generated by the leaky integrate and fire model. In addition, we discuss the similarity between the mapping mechanism of the present method and the structure of biological information processing from a point of view of self organization system theory.