MSc Stats Defence: Interpreting Capsule Networks for Classification by Routing Path Visualization

CANDIDATE:   AMAN BHULLAR

ABSTRACT:

Automating the classification of images is difficult in specialized areas because often there is insufficient labeled data to train a convolutional neural network, the most popular model for image classification [10]. Capsule networks, a neural network architecture proposed for image classification by Sabour et al. 2017, have been shown to require a smaller dataset than convolutional neural networks to train well [13]. In this thesis, a capsule network model that can classify astronomical images as containing or not containing at least one supernova light echo is identified, and shown to obtain an accuracy of 90% on the test set. In addition, routing path visualization, a technique for interpreting the entity that a given capsule in a capsule networks detects, is introduced in this thesis. Experimental results demonstrate that routing path visualization can also be used to precisely localize supernova light echoes in astronomical images.