Abstract:Deep neural networks (DNNs) on Riemannian manifolds have garnered increasing interest in various applied areas. For instance, DNNs on spherical and hyperbolic manifolds have been designed to solve a wide range of computer vision and nature language processing tasks. One of the key factors that contribute to the success of these networks is that spherical and hyperbolic manifolds have the rich algebraic structures of gyrogroups and gyrovector spaces. This enables principled and effective generalizations of the most successful DNNs to these manifolds. Recently, some works have shown that many concepts in the theory of gyrogroups and gyrovector spaces can also be generalized to matrix manifolds such as Symmetric Positive Definite (SPD) and Grassmann manifolds. As a result, some building blocks for SPD and Grassmann neural networks, e.g., isometric models and multinomial logistic regression (MLR) can be derived in a way that is fully analogous to their spherical and hyperbolic counterparts. Building upon these works, we design fully-connected (FC) and convolutional layers for SPD neural networks. We also develop MLR on Symmetric Positive Semi-definite (SPSD) manifolds, and propose a method for performing backpropagation with the Grassmann logarithmic map in the projector perspective. We demonstrate the effectiveness of the proposed approach in the human action recognition and node classification tasks.

What problem does this paper attempt to address?

This paper primarily discusses the application of deep neural networks (DNNs) on Riemannian manifolds, particularly on the symmetric positive definite (SPD) and Grassmann manifolds. The authors point out that despite the rich structure and computational feasibility provided by the SPD and Grassmann manifolds, existing neural network approaches lack the necessary techniques and mathematical tools to construct DNNs on these manifolds and cannot naturally generalize the basic components of DNNs as they do on the spherical and hyperbolic spaces. The main contributions of this paper include: 1. Generalizing fully connected (FC) and convolutional layers to the SPD manifold. 2. Proposing a method of using the Grassmann logarithmic mapping for backpropagation without approximation schemes and demonstrating how to construct graph convolutional networks (GCNs) on the Grassmann manifold. 3. Developing polynomial logistic regression (PLR) on the symmetric positive semidefinite (SPSD) manifold. 4. Demonstrating the effectiveness of the proposed methods in human action recognition and node classification tasks. The authors first review the fundamental concepts of the SPD and Grassmann manifolds, as well as existing work on constructing neural networks on these manifolds. They then propose a "hyperbolic vector space" approach that allows for the extension of FC and convolutional layer concepts to the SPD manifold and implements PLR on the SPSD manifold. Furthermore, they address the issue of using the Grassmann logarithmic mapping for backpropagation on the Grassmann manifold, which is crucial for training neural networks. In summary, the paper aims to expand the scope of deep learning by introducing new geometric operations and backpropagation strategies to overcome the limitations of constructing DNNs on the SPD and Grassmann manifolds.

Matrix Manifold Neural Networks++

Building Neural Networks on Matrix Manifolds: A Gyrovector Space Approach

Dynamically Stable Poincaré Embeddings for Neural Manifolds

Deep Manifold Computing and Visualization

Deep Manifold Transformation for Nonlinear Dimensionality Reduction

Deep Manifold Part 1: Anatomy of Neural Network Manifold

ManifoldNet: A Deep Network Framework for Manifold-valued Data

Spiking Graph Neural Network on Riemannian Manifolds

Exploring the Manifold of Neural Networks Using Diffusion Geometry

RMLR: Extending Multinomial Logistic Regression into General Geometries

Deep Manifold Learning of Symmetric Positive Definite Matrices with Application to Face Recognition.

Deep manifold-to-manifold transforming network for action recognition

Manifold GCN: Diffusion-based Convolutional Neural Network for Manifold-valued Graphs

Learning a discriminative SPD manifold neural network for image set classification

A Convergence Rate for Manifold Neural Networks

Manifold Kernel Sparse Representation of Symmetric Positive-Definite Matrices and Its Applications

Markov-Lipschitz Deep Learning

Riemannian Residual Neural Networks

ManifoldNorm: Extending normalizations on Riemannian Manifolds

A Riemannian Network for SPD Matrix Learning

Riemannian Local Mechanism for SPD Neural Networks