CNN LEARNING METHOD AND LEARNING DEVICE FOR ADJUSTING PARAMETERS OF CNN IN WHICH RESIDUAL NETWORKS ARE PROVIDED FOR META LEARNING AND TESTING METHOD AND TESTING DEVICE USING THE SAME

摘要

Meta Learning, that is, a convolutional layer that generates an output feature map by applying a convolution operation to an image or an input feature map corresponding thereto, in order to learn a learning method, and the convolutional layer or a subconvolutional layer corresponding thereto A Convolutional Neural Network (CNN)-based method using a learning device including a residual network that bypasses and feeds forwards the image or an input feature map corresponding thereto to the next convolutional layer is provided. The CNN-based method includes: (a) selecting a specific residual network to be dropped out from among the residual networks; (b) generating a CNN output by inputting the image to a modified CNN in which the specific residual network is dropped; and (c) calculating a loss by using the CNN output and a ground truth (GT) corresponding thereto, and adjusting the parameters of the modified CNN. In addition, the CNN-based method may be applied to layer-wise dropout, stochastic ensemble, and virtual driving.