帮我改成深度学习风格nIn-the-field-of-deep-learning-MLPs-are-once-again-gaining-attention-from-researchersThis-paper-introduces-MC-MLP-a-general-MLP-like-backbone-for-computer-vision-which-is-composed-of-a-series-of-fully-connected-FC-layersIn-MC-MLP-we-propose

In the realm of deep learning, MLPs have once again piqued the interest of researchers. This study presents MC-MLP, a general MLP-like architecture for computer vision, which is composed of several fully-connected (FC) layers. In MC-MLP, we propose that identical semantic information poses unique learning challenges in various coordinate frames. Specifically, we apply an orthogonal transformation to the feature information to alter the coordinate system, while simultaneously accomplishing the aggregation of feature information from diverse tokens. This design endows MC-MLP with multi-coordinate system receptive fields and the ability to adaptively learn information from different coordinate systems. The experiments demonstrate that MC-MLP outperforms most MLP models in image classification tasks and achieves superior performance at the same parameter level. The codes and models are publicly available at: https://github.com/MC-MLP.