작품개요

Since expensive equipment called GPUs is needed to learn deep learning models, computer resources called GPUs are limited to individuals in most situations, and it is a major problem to improve the model's performance as much as possible while efficiently utilizing a given limited GPU. Therefore, this study verifies that the efficient knowledge distortion (K.D) method of GPU memory and inductive bias are low, so that the vision transformer (ViT) architecture has the potential to improve performance as learning continues to be used to improve.