Learning Characteristics of Reverse Quaternion Neural Network

The purpose of this paper is to propose a new neuron structure for multilayer feedforward quaternion neural networks called Reverse Quaternion Neural Network, which utilizes the non-commutative property of quaternion multiplication, and to clarify its learning characteristics. While quaternion neural networks have been used in various fields, there have been no research reports on the characteristics of multi-layer feedforward quaternion neural networks that apply weights in the reverse direction. In this paper, we examined the learning characteristics of the Reverse Quaternion Neural Network through two experimental studies: line rotation tasks and three-dimensional object rotation tasks using ModelNet10. Through comparative analysis with conventional models, we investigated both the generalization capabilities and internal learning dynamics.Our results demonstrated that while the Reverse Quaternion Neural Network achieved performance comparable to existing models, it exhibited distinct characteristics in both rotational generalization capability and learning behavior, as evidenced by its gradient dynamics and feature acquisition process.