hZACH-ViT: Curved Latent Geometry for Compact Vision Transformers in Low-Data Medical Imaging
Researchers have developed hZACH-ViT, a new family of Vision Transformers designed for medical imaging in low-data environments. This model modifies the latent geometry of existing ZACH-ViT architectures, exploring non-Euclidean spaces like hyperbolic and spherical geometries instead of the standard Euclidean. Experiments on seven MedMNIST datasets showed that these curved latent geometries, particularly with low curvature, consistently improved performance over the Euclidean baseline, suggesting geometry is a dataset-dependent variable for model selection. AI
IMPACT Introduces curved latent geometry as a tunable parameter for improving vision transformer performance in low-data medical imaging tasks.