Co-Training 2L Submodels for Visual Recognition | Read Paper on Bytez