Transferability of Structure Parameters and Features in Convolutional Neural Network Layers
Xinhuan Luo · 2021
Many convolutional neural networks (CNNs) trained on images exhibit a phenomenon in common: the features extracted by a CNN's lower-level layers tend to be general, while higher-level layers extract more specific features that are limited to datasets and tasks. Based on this network characteristic, many scholars have demonstrated the transferability of convolutional network layers through experiments. In this study, we explored the transferability of each layer of the CNNs from two perspectives: network structure parameters and output features. First, the similarities of a network's structure parameters and output features were quantified using common distance measures (e.g., Euclidean distance). Then we conducted a transfer experiment to compare and analyze the relationship between the similarity and transferability of each layer of the network. We concluded that the more similar the structure parameters and output features of each layer of the CNN, the better the transfer effect.