[5M2-GS-2c] Machine learning

12:30 PM - 12:45 PM JST(3:30 AM - 3:45 AM UTC)

[5M2-GS-2c-03] Bayesian Snake Neural Network for Periodic Multivariate Time Series Analysis

〇Kota KURIHARA¹, Masahiro KOHJIMA², Moeka YOSHINARI², Ryuji YAMAMOTO², Yasuhiro MINAMI¹ (1. The University of Electro-Communications, 2. NTT)

Keywords:

Machine Learning,Bayesian Estimation,Time Series Data

This study proposes a method that enables high-precision analysis and uncertainty quantification using neural networks, specifically for data containing periodic components, such as physiological signals and environmental data (e.g., temperature, humidity, CO2 levels). The key to constructing our method lies in combining the snake function to capture periodic patterns with Bayesian estimation to provide prediction uncertainty (confidence). By utilizing Monte Carlo Dropout, our method realizes (approximate) Bayesian inference without modifying the training or inference processes of existing neural networks. We show the effectiveness of the proposed method through experiments using benchmark time-series data.