Our method for generalizing fully connected (FC) layers to the Symmetric Positive Definite (SPD) manifold setting relies on a reformulation of SPD hypergyroplanes, allowing the development of neural networks that can better handle structured data.
The interpretation of FC layers as distances from the output to hyperplanes orthonormal to specific axes in the output space has enabled advancements in adapting such layers to the hyperbolic setting.
#neural-networks #symmetric-positive-definite-manifolds #human-action-recognition #machine-learning #deep-learning
Collection
[
|
...
]