Inception and ResNet: Same Training, Same Features

摘要

Deep convolutional neural networks (CNNs) are the dominant technology in computer vision today. Unfortunately, it's not clear how different from each other the best CNNs really are. This paper measures the similarity between two well-known CNNs, Inception and ResNet, in terms of the features they extract from images. We find that Inception's features can be well approximated as an affine transformation of ResNet's features and vice-versa. The similarity between Inception and ResNet features is surprising. Convolutional neural networks learn complex non-linear features of images, and the architectural differences between the systems suggest that these functions should take different forms. Instead, they seem to have converged on similar solutions. This suggests that the selection of the training set may be more important than the selection of the convolutional architecture.