Feature Visualisation of Translation Invariance for Alexnet Network

Convolutional neural networks exhibit excellent performance in the image domain, and translation invariance is a desirable case for image recognition. Many studies have shown that deep convolutional networks can learn translation invariance through translation data augmentation. To help understand how deep convolutional networks achieve translation invariance through translational enhancement training, this paper uses the maximum activation response visualisation method for deep convolutional networks to visualise the features of Alexnet networks trained with translational enhancement. Through the feature visualisation experiments, it is found that the input images at different positions can excite different filters to make the maximum response. This shows that the deep convolutional network achieves translational invariance by setting the corresponding filters for each position separately to extract features. On the comparison experiments that were not trained with translation enhancement, the fully connected layer did not show significant translation invariance.