1: \begin{abstract}

2: Matrix manifolds, such as manifolds of Symmetric Positive Definite (SPD) matrices and Grassmann manifolds,

3: appear in many applications. 

4: %are commonly encountered in representation learning.  

5: Recently, by applying the theory of gyrogroups and gyrovector spaces that is a powerful framework for

6: studying hyperbolic geometry, 

7: some works have attempted to build principled generalizations of Euclidean neural networks on matrix manifolds. 

8: %This allows them to generalize some basic concepts in Euclidean spaces, 

9: %e.g. the binary operation, scalar multiplication, and function derivative to the matrix manifold setting in a principled way. 

10: However, due to the lack of many concepts in gyrovector spaces for the considered manifolds, 

11: e.g., the inner product and gyroangles,  

12: techniques and mathematical tools provided by these works are still limited compared to those developed 

13: for studying hyperbolic geometry.

14: In this paper, we generalize some notions in gyrovector spaces for SPD and Grassmann manifolds,  

15: and propose new models and layers for building neural networks on these manifolds. 

16: We show the effectiveness of our approach in two applications, i.e., human action recognition and knowledge graph completion.

17: \end{abstract}

18: