abstract |
Methods and systems using improved training and learning for deep neural networks are disclosed. In one example, a deep neural network includes multiple layers, and each layer has multiple nodes. For each L layer of the plurality of layers, the nodes of each L layer are randomly connected to nodes in the L+1 layer. For each L+1 layer in the plurality of layers, nodes in each L+1 layer are connected to nodes in subsequent L layers in a one-to-one manner. The parameters associated with the nodes of each L layer are fixed. Update the parameters related to the nodes of each L+1 layer, and L is an integer starting from 1. In another example, a deep neural network includes an input layer, an output layer, and multiple hidden layers. The labels of the input and output layers of the input layer are determined in relation to the first sample. A Gaussian regression procedure is used to estimate the similarity between different pairs of inputs and labels between the second sample and the first sample. |