Abstract:This paper presents a hybrid morphological neural network for regression tasks called linear dilation-erosion regressor ($\ell$-DER). An $\ell$-DER is given by a convex combination of the composition of linear and morphological operators. They yield continuous piecewise linear functions and, thus, are universal approximators. Besides introducing the $\ell$-DER model, we formulate their training as a difference of convex (DC) programming problem. Precisely, an $\ell$-DER is trained by minimizing the least-squares using the convex-concave procedure (CCP). Computational experiments using several regression tasks confirm the efficacy of the proposed regressor, outperforming other hybrid morphological models and state-of-the-art approaches such as the multilayer perceptron network and the radial-basis support vector regressor.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to effectively train a hybrid morphological neural network for regression tasks, namely the linear dilation - erosion perceptron (ℓ - DER). Specifically, the author proposes a new method to optimize the parameters of the ℓ - DER model by using the convex - concave procedure (CCP) to minimize the mean squared error (MSE). This method aims to overcome the challenges brought by the non - differentiability of the maximum and minimum operations in the training of traditional morphological neural networks. ### Main contributions of the paper: 1. **Propose a new regression model**: The linear dilation - erosion regressor (ℓ - DER), which combines linear transformation and morphological operators. 2. **Solve the training problem**: By transforming the training problem into a difference of convex (DC) optimization problem and using the CCP method to solve it, the non - differentiability problem of morphological operators is effectively dealt with. 3. **Verify the effectiveness of the model**: Through experiments on multiple regression tasks, it is proved that the ℓ - DER model is superior in performance to other hybrid morphological models as well as traditional methods such as the multi - layer perceptron (MLP) and support vector regression (SVR). ### Formula summary: - The decision function of ℓ - DER is: \[ \tau_\ell(x) = \delta_a(Wx) - \delta_b(Mx) \] where $\delta_a$ and $\delta_b$ represent the dilation and erosion operations respectively, $W$ and $M$ are weight matrices, and $a$ and $b$ are bias vectors. - The mean squared error (MSE) is defined as: \[ \text{MSE}(T, \alpha) = \frac{1}{m} \sum_{i = 1}^m (y_i - \tau_\ell(x_i; \alpha))^2 \] where $T=\{(x_i, y_i):i = 1,\ldots,m\}$ is the training set and $\alpha$ is the model parameter. - The optimization problem can be written as: \[ \minimize_{\alpha} \frac{1}{m} \sum_{i = 1}^m (y_i - \tau_\ell(x_i; \alpha))^2 \] Through these methods, the paper successfully proposes and verifies a new and efficient regression model and its training method.

Least-Squares Linear Dilation-Erosion Regressor Trained using a Convex-Concave Procedure

Linear Dilation-Erosion Perceptron for Binary Classification

Wordreg: Mitigating the Gap Between Training and Inference with Worst-Case Drop Regularization

Training Single-Layer Morphological Perceptron Using Convex-Concave Programming

Nonlinear Regression via Deep Negative Correlation Learning

Relaxed Earth Mover's Distances for Chain- and Tree-connected Spaces and their use as a Loss Function in Deep Learning

Derivative-based regularization for regression

Advances in the training, pruning and enforcement of shape constraints of Morphological Neural Networks using Tropical Algebra

Learning Smooth Neural Functions via Lipschitz Regularization

Learning Deep Morphological Networks with Neural Architecture Search

Diffeomorphic Latent Neural Operators for Data-Efficient Learning of Solutions to Partial Differential Equations

Deep Multi-View Learning using Neuron-Wise Correlation-Maximizing Regularizers

A Neural-Network-Based Convex Regularizer for Inverse Problems

DELR-Net: a network for 3D multimodal medical image registration in more lightweight application scenarios

A derived least square extreme learning machine

Deep Loss Convexification for Learning Iterative Models

Neural ePDOs: Spatially Adaptive Equivariant Partial Differential Operator Based Networks

Learning Homeomorphic Image Registration via Conformal-Invariant Hyperelastic Regularisation

Error-Correcting Neural Networks for Two-Dimensional Curvature Computation in the Level-set Method

Learning Euler's Elastica Model for Medical Image Segmentation

DenseReg: Fully Convolutional Dense Shape Regression In-the-Wild