Mobile V-MoEs: Scaling Down Vision Transformers via Sparse Mixture-of-Experts