Analysis and interpretation of Deep Convolutional Features using Self-Organizing Maps

Abstract

Deep learning has defined a new paradigm for data analysis. In image processing, Convolutional Neural Networks (CNN) have a vast number of appli-cations and do not require prior extraction of features as these are “learned” di-rectly from training images. The interpretation about how a CNN works is an open problem and any method for interpreting the features extracted from CNN can lead to remove the black-box concept significant contribution to the field machine-learning. In the present chapter, a SOM-based approach for analysis and interpretation of features extracted from CNN is proposed. Main characteristics are: i) CNN are trained from an initial image dataset with different sets of hy-perparameters; ii) new datasets containing different representations of the initial dataset are generated and then analyzed using SOM, visualization tools, and qual-ity measures; iii) it is possible to select features suitable for classification and to describe complexity and diversity in the classes and to extract additional infor-mation about the images in the training datasets. An application example consid-ering chest X-ray images for classification of pneumonia is analyzed, being to identify good features from CNN from scratch and to give some interpretation from them both in the classification of normal vs pneumonia, and in viral pneu-monia vs bacterial pneumonia.