Necessity of Increasing Kernel Size to Secure Receptive Fields in CNN for Time Series Analysis

Convolutional neural network (CNN) is widely used for analyzing time series data as it allows for the rapid learning of inherent characteristics in the series with a small number of parameters through filter operations. To prevent overfitting while maintaining a small kernel size and increasing the receptive field, dilated convolution has been proposed and effectively applied in the field of computer vision and time series. However, dilated convolution has gaps within the kernel, making it ineffective at capturing spectral information. We demonstrate through sim-ulations and real electroencephalogram (EEG) data that neural signals can be more effectively analyzed by directly increasing the kernel size instead of using dilated convolution. Our experimental results show that directly increasing the kernel size according to the sampling rate and the frequency bands of interest is crucial.