The Geometric Revolution That's Making Computer Vision More Efficient

This work addresses the instability and computational expense of existing approaches by presenting an improved framework for hyperbolic deep learning suited to computer vision problems. In addition to a new Riemannian AdamW optimizer that immediately adapts to changing manifold geometries, we present a novel curvature-learning technique. Furthermore, even with decreased numerical precision, our convolutional and normalizing enhancements enable effective large-scale hyperbolic modeling. The suggested method minimizes projection errors, stabilizes curvature adaption, and scales well to high-dimensional data. Significant performance improvements are shown in experimental results for both classification and hierarchical metric learning, opening the door for more adaptable and scalable hyperbolic vision architectures.