🏢 Kiel University

HydraViT: Stacking attention heads creates a scalable Vision Transformer, adapting to diverse hardware by dynamically selecting subnetworks during inference, improving accuracy and efficiency.