Abstract:Real-world 3D structured data like point clouds and skeletons often can be represented as data in a 3D rotation group (denoted as $\mathbb{SO}(3)$). However, most existing neural networks are tailored for the data in the Euclidean space, which makes the 3D rotation data not closed under their algebraic operations and leads to sub-optimal performance in 3D-related learning tasks. To resolve the issues caused by the above mismatching between data and model, we propose a novel non-real neuron model called \textit{quaternion product unit} (QPU) to represent data on 3D rotation groups. The proposed QPU leverages quaternion algebra and the law of the 3D rotation group, representing 3D rotation data as quaternions and merging them via a weighted chain of Hamilton products. We demonstrate that the QPU mathematically maintains the $\mathbb{SO}(3)$ structure of the 3D rotation data during the inference process and disentangles the 3D representations into ``rotation-invariant'' features and ``rotation-equivariant'' features, respectively. Moreover, we design a fast QPU to accelerate the computation of QPU. The fast QPU applies a tree-structured data indexing process, and accordingly, leverages the power of parallel computing, which reduces the computational complexity of QPU in a single thread from $\mathcal{O}(N)$ to $\mathcal {O}(\log N)$. Taking the fast QPU as a basic module, we develop a series of quaternion neural networks (QNNs), including quaternion multi-layer perceptron (QMLP), quaternion message passing (QMP), and so on. In addition, we make the QNNs compatible with conventional real-valued neural networks and applicable for both skeletons and point clouds. Experiments on synthetic and real-world 3D tasks show that the QNNs based on our fast QPUs are superior to state-of-the-art real-valued models, especially in the scenarios requiring the robustness to random rotations.

Neural network models for the quaternion singular value decompositions

Optimizing Quantized Neural Networks in a Weak Curvature Manifold

A Quaternion-Valued Neural Network Approach to Nonsmooth Nonconvex Constrained Optimization in Quaternion Domain

Tensor Neural Network Models for Tensor Singular Value Decompositions.

Quaternion Convolutional Neural Networks

A novel structure-preserving algorithm for the singular value decomposition of biquaternion matrices

Fast Quaternion Product Units for Learning Disentangled Representations in SO(3)

Quaternion Product Units for Deep Learning on 3D Rotation Groups

On Quaternion Higher-Order Singular Value Decomposition: Models and Analysis

On singular value decomposition and generalized inverse of a commutative quaternion matrix and applications

Quaternion Matrix Optimization and The Underlying Calculus

Quaternion Space Sparse Decomposition for Motion Compression and Retrieval

Stability Analysis of Quaternion-Valued Neural Networks: Decomposition and Direct Approaches.

Dequantizing quantum machine learning models using tensor networks

Singular Value Decomposition of Third Order Quaternion Tensors

Global stability and synchronization of stochastic discrete-time variable-order fractional-order delayed quaternion-valued neural networks

Dual complex structure-preserving algorithm of dual quaternion singular value decomposition and its applications

Neural Network for Computing GSVD and RSVD

Quaternion Singular Value Decomposition Approach to Color Image De-Noising

Learning Algorithms in Quaternion Neural Networks Using Ghr Calculus

Quaternion tensor singular value decomposition using a flexible transform-based approach