Abstract:We develop a unified model, known as MgNet, that simultaneously recovers some convolutional neural networks (CNN) for image classification and multigrid (MG) methods for solving discretized partial differential equations (PDEs). This model is based on close connections that we have observed and uncovered between the CNN and MG methodologies. For example, pooling operation and feature extraction in CNN correspond directly to restriction operation and iterative smoothers in MG, respectively. As the solution space is often the dual of the data space in PDEs, the analogous concept of feature space and data space (which are dual to each other) is introduced in CNN. With such connections and new concept in the unified model, the function of various convolution operations and pooling used in CNN can be better understood. As a result, modified CNN models (with fewer weights and hyper parameters) are developed that exhibit competitive and sometimes better performance in comparison with existing CNN models when applied to both CIFAR-10 and CIFAR-100 data sets.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to simultaneously recover the convolutional neural network (CNN) for image classification and the multigrid method (MG) for solving discrete partial differential equations (PDEs) by establishing a unified model (called MgNet). This model is based on the close connections between CNN and MG methods observed and discovered by the author. For example, the pooling operation and feature extraction in CNN directly correspond to the restriction operation and iterative smoother in MG respectively. In addition, since the solution space of PDEs is usually the dual of the data space, a similar concept - feature space and data space (which are dual) - is introduced in CNN. Through these connections and new concepts, the functions of various convolutional operations and pooling in CNN can be better understood. As a result, improved CNN models with fewer weights and hyper - parameters are developed, and these models exhibit performance comparable to or even better than existing CNN models when applied to the CIFAR - 10 and CIFAR - 100 datasets. Specifically, the main contributions of the paper are as follows: 1. **Establishing the connection between CNN and MG**: The paper elaborately explores the similarities and differences between CNN and MG methods, especially how to map the concepts in MG to CNN. 2. **Proposing the MgNet model**: By combining the methods of CNN and MG, a new neural network structure MgNet is proposed, and this model can achieve excellent performance in image classification tasks. 3. **Theoretical analysis**: The theoretical framework of MgNet is provided, explaining why this combination can be effective and showing the relationship between MgNet and existing models such as ResNet. 4. **Experimental verification**: Through experiments on the CIFAR - 10 and CIFAR - 100 datasets, the effectiveness and superiority of MgNet are proved. In conclusion, this paper aims to provide a new perspective to understand and improve deep - learning models, especially in the application of image classification tasks, by building a bridge between CNN and MG.

MgNet: A Unified Framework of Multigrid and Convolutional Neural Network

Multinoise-type Blind Denoising Using a Single Uniform Deep Convolutional Neural Network.

GCFnet: Global Collaborative Fusion Network for Multispectral and Panchromatic Image Classification

MgNO: Efficient Parameterization of Linear Operators via Multigrid

Meta-MgNet: Meta Multigrid Networks for Solving Parameterized Partial Differential Equations.

MGCNN: a learnable multigrid solver for sparse linear systems from PDEs on structured grids

MGFN: A Multi-Granularity Fusion Convolutional Neural Network for Remote Sensing Scene Classification

A Multigrid Graph U-Net Framework for Simulating Multiphase Flow in Heterogeneous Porous Media

Multi-Glimpse Network: A Robust and Efficient Classification Architecture based on Recurrent Downsampled Attention.

DiffGCN: Graph Convolutional Networks via Differential Operators and Algebraic Multigrid Pooling

MGMNet: Mutual-Guidance Mechanism for Joint Classification of Multisource Remote Sensing Data

Hybrid Convolutional Network Combining Multiscale 3D Depthwise Separable Convolution and CBAM Residual Dilated Convolution for Hyperspectral Image Classification

Multi-scale Unified Network for Image Classification

CNN-GCN Joint Network for Remote Sensing Scene Classification

Multiresolution Interpretable Contourlet Graph Network for Image Classification

Mutual Guidance and Residual Integration for Image Enhancement

CMPF-UNet: a ConvNeXt multi-scale pyramid fusion U-shaped network for multi-category segmentation of remote sensing images

HDECGCN: A Heterogeneous Dual Enhanced Network Based on Hybrid CNNs Joint Multiscale Dynamic GCNs for Hyperspectral Image Classification

Multiscale Random-Shape Convolution and Adaptive Graph Convolution Fusion Network for Hyperspectral Image Classification

Hyper-LGNet: Coupling Local and Global Features for Hyperspectral Image Classification

MG-Net: Multi-level global-aware network for thymoma segmentation