Work

Understanding convolutional neural networks through Z-transform: Frequency domain analysis of kernel feature extraction

Year: 2025

Type: article

Abstract: This paper analyzes Convolutional Neural Networks (CNNs) from a spatial frequency perspective using Z-transforms to evaluate kernel transfer functions. By examining which frequency bands kernels respond to, we quantitatively analyze feature processing through CNN layers. Experiments with custom and pre-trained models (InceptionV3, VGG16, ResNet-50, DenseNet-121) reveal CNNs capture high-frequency features in early layers while increasingly focusing on low-frequency features in deeper layers. Kernel pruning experiments demonstrate that kernels with larger weights capturing low-frequency features are critical for model performance. These findings provide insights into CNN feature extraction processes, contributing to improved interpretability and model construction guidelines.

Source: Nonlinear Theory and Its Applications IEICE

DOI: 10.1587/nolta.16.878

Authors: Sora Togawa, Kenya Jin'no

Institution: Tokyo City University

Language: English

Topic: Advanced Neural Network Applications

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: diamond

Funders: Japan Society for the Promotion of Science, Research Institute of Electrical Communication, Tohoku University

Awards: トピックモデルと時空間深層学習を用いたトラフィック予測法に関する研究, Analysis of latent variables in deep learning and their compositional methods, Optimization of deep learning models using evolutionary computation methods based on analog circuits