HOLMES: HOLonym-MEronym based Semantic inspection

In recent years, Deep Learning models have improved the state of the art in several domains. In particular, Convolutional Neural Networks have arguably become the go-to solution for Computer Vision, thanks to their ability to automatize the feature extraction process in image classification tasks. However, the drawback of these models is that they are black-boxes: that is, the knowledge acquired in training is fully sub-symbolic and thus these models retain no human-understandable representation of their inner decision processes. There is an entire research field working towards describing the rationale behind AI decision-making, called Explainable AI (XAI). However, when it comes to Computer Vision, what all XAI techniques typically produce are heatmaps that measure the correlation of each input-pixel to the classification label. We argue that these shallow explanations are not enough for a human user to fully trust the algorithmic decision, nor for a developer to sufficiently debug a model in order to assess its learning progress. Conversely, when asked to justify an image-classification task, humans typically produce part-based explanations, e.g. ‘this image depicts a cat, because there are pointy ears up there and a tail there, etc.’ This thesis introduces HOLMES (HOLonym-MEronym based Semantic inspection), an algorithmic pipeline that, given a trained convolutional neural network and an image to be classified, automatically retrieves meronyms (parts) of the predicted label and produces several heatmaps at meronym level, thus highlighting the specific locations and importance of the parts that compose the main object. We argue that this is a fundamental step towards structured, ontology-linked explanations, enabling a deeper understanding of convolutional neural networks' learned knowledge and inner decision process.