CHEER: Rich Model Helps Poor Model via Knowledge Infusion