We introduce a data-dependent weight initialization scheme for ReLU and output layers commonly found in modern neural network architectures. An initial feedforward pass through the network is performed using an initialization set (a subset of the training data set). Using statistics obtained from this pass, we initialize the weights of the network, so the following properties are met: (1) weight matrices are orthogonal; (2) ReLU layers produce a predetermined fraction of nonzero activations; (3) the outputs produced by internal layers have a predetermined variance; (4) weights in the last layer are chosen to minimize the squared error in the initialization set. We evaluate our method on popular architectures (VGG16, VGG19, and InceptionV3) and faster convergence rates are achieved on the ImageNet data set when compared to state-of-the-art initialization techniques (LSUV, He, and Glorot).
展开▼
