METHOD, APPARATUS AND DEVICE FOR UPDATING CONVOLUTIONAL NEURAL NETWORK USING GPU CLUSTER

摘要

A method, apparatus and device for updating a convolutional neural network using a GPU cluster. The GPU cluster comprises a first GPU and several other GPUs. The method is executed by the first GPU, and comprises: obtaining a sample having a classification tag (S202); executing, on the basis of parameters of layers of networks at a front end, a first operation on the sample to obtain a first operation result (S204); executing, on the basis of the first operation result and comprised parameters of layers of networks at a back end, a second operation on the sample to obtain a second operation result (S206); respectively sending the first operation to other GPUs so that other GPUs respectively perform corresponding operations on the sample on the basis of respective comprised parameters of layers of networks at a back end, and the first operation result (S208); receiving a third operation result obtained by performing a corresponding third operation by other GPUs (S210); combining the second operation result with the third operation result to obtain the classification result (S212); determining, according to the classification result and the classification tag, a prediction error (S214); and updating the convolutional neural network according to the prediction error (S216).