This application provides a gradient transmission method and a related apparatus. In this method, a first communication apparatus receives training data, and determines a first intermediate gradient based on the training data. The first intermediate gradient is used to update a parameter of a first neural network located in a second communication apparatus. The first communication apparatus maps the first intermediate gradient to an air interface resource to generate a first gradient signal, and sends the first gradient signal to the second communication apparatus. The first gradient signal includes one or more first gradient symbols, and each of the first gradient symbols is corresponding to one or more gradient values. The first communication apparatus maps the first intermediate gradient to the air interface resource for transmission, that is, implements reverse transmission of the first intermediate gradient by sending the first gradient signal to the second communication apparatus, to facilitate implementing joint training of the first communication apparatus and the second communication apparatus.