Unsupervised clustering based understanding of CNN

Convolutional Neural networks have been very successful for most computer vision tasks such as image recognition, classification, object detection and segmentation. Even though CNNs are very successful and give superior results as compared to traditional image processing algorithms, interpretability of their results remains an important issue to be solved. Indeed, lack of interpretability and explainability of how CNN work at their various levels, caused a certain skepticism among their potential users, as for example those working in medical diagnosis or autonomous driving cars. The current study aims to answer some of the issues related to interpretability by the use unsupervised methods to discern the features learned by the CNN in different layers.