ManifoldNet: A Deep Neural Network for Manifold-Valued Data With Applications.

Area of Science:

• Geometric deep learning
• Manifold learning
• Computer vision
• Medical imaging

Background:

• Deep learning models typically process data in Euclidean spaces.
• Data from domains like computer vision (e.g., pose data) and medical imaging (e.g., diffusion tensors) often reside on non-Euclidean manifolds.
• Processing this manifold-valued data requires specialized architectures beyond standard deep learning.

Purpose of the Study:

• To develop a novel theoretical framework and architecture for deep neural networks capable of handling grids of manifold-valued data.
• To introduce ManifoldNet, an architecture designed for geometric deep learning on non-Euclidean data.
• To ensure that data remains manifold-valued throughout the network's hidden layers.

Main Methods:

• The proposed ManifoldNet architecture uses weighted Fréchet means for 'convolutions' on manifolds.
• A provably convergent recursive algorithm is presented to efficiently compute the Fréchet means.
• Theoretical analysis demonstrates that ManifoldNet layers act as contraction mappings and are equivariant to isometries on Riemannian manifolds.

Main Results:

• ManifoldNet successfully processes manifold-valued data, maintaining data type through hidden layers.
• The Fréchet mean computation is optimized for reduced computational complexity.
• The architecture exhibits properties crucial for deep networks: non-collapsibility and equivariance to group actions.

Conclusions:

• ManifoldNet provides a robust theoretical and architectural solution for geometric deep learning on manifold-valued data.
• The framework effectively handles complex data structures found in computer vision and medical imaging.
• Experimental results validate the performance and applicability of ManifoldNet on diverse datasets.