Optimal label vector for convolutional neural network
Lin Feng, Muxin Sun, Shenglan Liu, Jun Wu · 2016
Convolutional Neural Network(CNN) is a kind of deep learning and it has become a current hot topic in the field of image recognition. In the CNN, Output layer consists of Euclidean Radial Basis Function, unit matrix column as CNN's label vector. The category of the input image can be interpreted as the nearest label vector. This paper addresses a question: what is CNN's optimal label vector? We show that label vector influence CNN's accuracy. Most surprisingly, we show that the new label vector achieves the lowest known error rate on the Convolutional net LeNet-5, unprocessed MNIST dataset (0.45%).