Wavelet Convolutions for Large Receptive Fields

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Finder, Shahaf E., Amoyal, Roy, Treister, Eran, Freifeld, Oren
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866929421157400576
author Finder, Shahaf E.
Amoyal, Roy
Treister, Eran
Freifeld, Oren
author_facet Finder, Shahaf E.
Amoyal, Roy
Treister, Eran
Freifeld, Oren
contents In recent years, there have been attempts to increase the kernel size of Convolutional Neural Nets (CNNs) to mimic the global receptive field of Vision Transformers' (ViTs) self-attention blocks. That approach, however, quickly hit an upper bound and saturated way before achieving a global receptive field. In this work, we demonstrate that by leveraging the Wavelet Transform (WT), it is, in fact, possible to obtain very large receptive fields without suffering from over-parameterization, e.g., for a $k \times k$ receptive field, the number of trainable parameters in the proposed method grows only logarithmically with $k$. The proposed layer, named WTConv, can be used as a drop-in replacement in existing architectures, results in an effective multi-frequency response, and scales gracefully with the size of the receptive field. We demonstrate the effectiveness of the WTConv layer within ConvNeXt and MobileNetV2 architectures for image classification, as well as backbones for downstream tasks, and show it yields additional properties such as robustness to image corruption and an increased response to shapes over textures. Our code is available at https://github.com/BGU-CS-VIL/WTConv.
format Preprint
id arxiv_https___arxiv_org_abs_2407_05848
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Wavelet Convolutions for Large Receptive Fields
Finder, Shahaf E.
Amoyal, Roy
Treister, Eran
Freifeld, Oren
Computer Vision and Pattern Recognition
In recent years, there have been attempts to increase the kernel size of Convolutional Neural Nets (CNNs) to mimic the global receptive field of Vision Transformers' (ViTs) self-attention blocks. That approach, however, quickly hit an upper bound and saturated way before achieving a global receptive field. In this work, we demonstrate that by leveraging the Wavelet Transform (WT), it is, in fact, possible to obtain very large receptive fields without suffering from over-parameterization, e.g., for a $k \times k$ receptive field, the number of trainable parameters in the proposed method grows only logarithmically with $k$. The proposed layer, named WTConv, can be used as a drop-in replacement in existing architectures, results in an effective multi-frequency response, and scales gracefully with the size of the receptive field. We demonstrate the effectiveness of the WTConv layer within ConvNeXt and MobileNetV2 architectures for image classification, as well as backbones for downstream tasks, and show it yields additional properties such as robustness to image corruption and an increased response to shapes over textures. Our code is available at https://github.com/BGU-CS-VIL/WTConv.
title Wavelet Convolutions for Large Receptive Fields
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2407.05848