Multiscale Training of Convolutional Neural Networks

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ahamed, Shadab, Zakariaei, Niloufar, Haber, Eldad, Eliasof, Moshe
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914358045442048
author Ahamed, Shadab
Zakariaei, Niloufar
Haber, Eldad
Eliasof, Moshe
author_facet Ahamed, Shadab
Zakariaei, Niloufar
Haber, Eldad
Eliasof, Moshe
contents Training convolutional neural networks (CNNs) on high-resolution images is often bottlenecked by the cost of evaluating gradients of the loss on the finest spatial mesh. To address this, we propose Multiscale Gradient Estimation (MGE), a Multilevel Monte Carlo-inspired estimator that expresses the expected gradient on the finest mesh as a telescopic sum of gradients computed on progressively coarser meshes. By assigning larger batches to the cheaper coarse levels, MGE achieves the same variance as single-scale stochastic gradient estimation while reducing the number of fine mesh convolutions by a factor of 4 with each downsampling. We further embed MGE within a Full-Multiscale training algorithm that solves the learning problem on coarse meshes first and "hot-starts" the next finer level, cutting the required fine mesh iterations by an additional order of magnitude. Extensive experiments on image denoising, deblurring, inpainting and super-resolution tasks using UNet, ResNet and ESPCN backbones confirm the practical benefits: Full-Multiscale reduces the computation costs by 4-16x with no significant loss in performance. Together, MGE and Full-Multiscale offer a principled, architecture-agnostic route to accelerate CNN training on high-resolution data without sacrificing accuracy, and they can be combined with other variance-reduction or learning-rate schedules to further enhance scalability.
format Preprint
id arxiv_https___arxiv_org_abs_2501_12739
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Multiscale Training of Convolutional Neural Networks
Ahamed, Shadab
Zakariaei, Niloufar
Haber, Eldad
Eliasof, Moshe
Machine Learning
Training convolutional neural networks (CNNs) on high-resolution images is often bottlenecked by the cost of evaluating gradients of the loss on the finest spatial mesh. To address this, we propose Multiscale Gradient Estimation (MGE), a Multilevel Monte Carlo-inspired estimator that expresses the expected gradient on the finest mesh as a telescopic sum of gradients computed on progressively coarser meshes. By assigning larger batches to the cheaper coarse levels, MGE achieves the same variance as single-scale stochastic gradient estimation while reducing the number of fine mesh convolutions by a factor of 4 with each downsampling. We further embed MGE within a Full-Multiscale training algorithm that solves the learning problem on coarse meshes first and "hot-starts" the next finer level, cutting the required fine mesh iterations by an additional order of magnitude. Extensive experiments on image denoising, deblurring, inpainting and super-resolution tasks using UNet, ResNet and ESPCN backbones confirm the practical benefits: Full-Multiscale reduces the computation costs by 4-16x with no significant loss in performance. Together, MGE and Full-Multiscale offer a principled, architecture-agnostic route to accelerate CNN training on high-resolution data without sacrificing accuracy, and they can be combined with other variance-reduction or learning-rate schedules to further enhance scalability.
title Multiscale Training of Convolutional Neural Networks
topic Machine Learning
url https://arxiv.org/abs/2501.12739