Improving Deep Learning with Lorentzian Geometry: Results from LHIER Experiments

This study proposes a whole set of enhancements for hyperbolic deep learning in computer vision, which have been verified by conducting extensive experiments on conventional classification tasks and hierarchical metric learning. An effective convolutional layer, a resilient curvature learning schema, maximum distance rescaling for numerical stability, and a Riemannian AdamW optimizer are among the suggested techniques that are included into a Lorentz-based model (LHIER+). With greater Recall@K scores, LHIER+ performs better on hierarchical metric learning benchmarks (CUB, Cars, SOP).