Study on Characteristics of Capsule Network based on Affine Transformation

In view of the problem of spatial information defects expressed in Convolution Neural Network (CNN), G. E. Hinton, one of the founders of deep learning, proposed the Capsule Network Model. Compared with that of CNNs, the feature of the Capsule model takes the form of vectors, which convey the location transformation information from part to whole. In order to better understand the characteristics of the capsule network under the basic affine transformation, this paper uses Tensorboard that is an integrated simulation tool in Tensorflow, from the perspective of visualization, to study the characteristics of the capsule network under the basic affine transformation, such as translation and rotation. In addition, this paper provides an evaluation index for the accuracy of experimental results during translation and rotation. The experimental results show that the Capsule Network can actually learn and output "instantiation parameters" including the correct attitude and deformation information of the visual entity through the internal integrated capsule module. The Capsule Network is more intuitive to deal with the position and direction changes than the traditional neural networks.