Abstract:We propose a new approach, called as functional deep neural network (FDNN), for classifying multi-dimensional functional data. Specifically, a deep neural network is trained based on the principle components of the training data which shall be used to predict the class label of a future data function. Unlike the popular functional discriminant analysis approaches which rely on Gaussian assumption, the proposed FDNN approach applies to general non-Gaussian multi-dimensional functional data. Moreover, when the log density ratio possesses a locally connected functional modular structure, we show that FDNN achieves minimax optimality. The superiority of our approach is demonstrated through both simulated and real-world datasets.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that in multi - dimensional functional data classification, existing methods usually rely on the Gaussian assumption, which is often violated in practical applications, resulting in poor classification performance. The paper proposes a new method - Functional Deep Neural Network (FDNN) for handling the classification problem of general non - Gaussian multi - dimensional functional data. Specifically, the paper extracts the functional principal components of data functions through Functional Principal Component Analysis (FPCA), and then trains a Deep Neural Network (DNN) based on these principal components and their corresponding class labels. This method is not only applicable to the classification of complex curves or imaging data, but also has good theoretical properties. In particular, when the log - density ratio has a functionally - modular structure with local connections, FDNN can achieve minimax optimality. ### Main contributions of the paper: 1. **Theoretical contribution**: The paper establishes the sharp convergence rate of Minimax Excess Misclassification Risk (MEMR) when the data is of functional type, and this result can be applied to a wide range of functional data with complex density functions. 2. **Methodological contribution**: The proposed FDNN classifier can handle various one - dimensional or multi - dimensional functional data, and especially in the case of non - Gaussian data, it performs better than traditional classification methods. ### Specific content of the paper: - **Background and motivation**: The development of modern technology has made complex functional data common, while classical multivariate analysis techniques such as logistic regression or discriminant analysis are no longer applicable to functional data because they are essentially infinite - dimensional. Existing methods based on Functional Principal Component Analysis (FPCA), such as functional discriminant analysis, usually assume that the data is a Gaussian process, which is often not true in practice. - **Method introduction**: The paper proposes the FDNN classifier. First, it extracts the functional principal components of data functions through FPCA, and then trains a DNN based on these principal components. The paper proves that when the log - density ratio has a functionally - modular structure with local connections, FDNN can achieve minimax optimality. - **Experimental verification**: The effectiveness of FDNN is verified through simulated data and real - world data sets (such as speech recognition data and Alzheimer's disease data). Experimental results show that FDNN outperforms existing classification methods in multiple situations, especially when dealing with non - Gaussian data. ### Key terms: - **Functional classification** - **Functional data analysis** - **Functional neural networks** - **Minimax excess misclassification risk** - **Multi - dimensional functional data** ### Conclusion: The Functional Deep Neural Network (FDNN) classifier proposed in the paper performs excellently when dealing with non - Gaussian complex functional data. It not only achieves minimax optimality theoretically, but also shows superior performance in practical applications.

Deep Neural Network Classifier for Multi-dimensional Functional Data

Multiclass classification for multidimensional functional data through deep neural networks

Functional data analysis using deep neural networks

Neural Networks as Functional Classifiers

Optimal Linear Discriminant Analysis for High-Dimensional Functional Data

Deep Learning With Functional Inputs

SFMD‐X: A New Functional Data Classifier Based on Shrinkage Functional Mahalanobis Distance

Nonlinear functional regression by functional deep neural network with kernel embedding

Functional Neural Networks for High-Dimensional Genetic Data Analysis

Functional Network: A Novel Framework for Interpretability of Deep Neural Networks

Robust Deep Neural Network Estimation for Multi-dimensional Functional Data

Multimodal functional deep learning for multiomics data

Deep Learning for Functional Data Analysis with Adaptive Basis Layers

Functional Sufficient Dimension Reduction for Functional Data Classification

HCFNN: High-order Coverage Function Neural Network for Image Classification

Complex Functional Networks

SD-CNN: A Static-Dynamic Convolutional Neural Network for Functional Brain Networks

Optimal Bayes Classifiers for Functional Data and Density Ratios

A Hierarchical Fused Fuzzy Deep Neural Network for Data Classification

Model-Based Clustering and Classification of Functional Data

Enriched Functional Tree-Based Classifiers: A Novel Approach Leveraging Derivatives and Geometric Features