METHOD FOR REDUCING THE SIZE OF AN ARTIFICIAL NEURAL NETWORK

摘要

Method of reducing the size of a teaching neural network to obtain a student neural network, comprising the steps of: - training (201) the teaching network to solve a first given problem for a set of training data, - During the training of the teaching network, i. Extract the output characteristics of the teacher network layer as a vector having a first dimension p l , ii. Apply (202) a dimension reduction technique to feature vectors to generate a projection matrix in a subspace of dimension k l strictly less than the first dimension p l , - Define the size of each intermediate layer of the student network as equal to the k l dimension of the subspace, - Train (203) the student network on the same set of training data for jointly solve the first problem and a second problem of reconstruction of the characteristics extracted from the teaching network projected in the subspaces of dimensions k l .