Dear all,
you are cordially invited to join the master's thesis
defense of Patrick Hilme on 15.12.22 at 14:00 in room 05.013 (CogSys Lab).
Patrick Hilme will defend his thesis Verbally Explaining Image Classications from Relevant Visual Features: Towards Learning Symbolic Structures
Using Layer-wise Relevance Propagation and Logic Programming
Abstract:
Concept Relevance Propagation (CRP) is an extension to Layer-wise Relevance Propagation (LRP). It is able to identify visual features in latent componentsof a DNN, that contribute to predictions, enabling visual explanations both locally and globally. This thesis merges the visual explanatory power of CRP with relation learning driven by Inductive Logic Programming (ILP) to generate verbal explanations for an image classifying network. Using CRP,visual features are discovered, transformed into a geometrical representation, before they are put into topological and spatial context to each other. Their identied relational structure is then formalized into rst-order logic. With this general procedure, relational learning is applied to contrastive image data with the ILP Framework Aleph to construct a symbolic hypothesis of the prediction. With visual features being symbolised in such a way it is argued, that
it can help to supply contrastive explanations in the form of near misses verbally. Furthermore, this thesis explores, if this symbolic approach can help to
discover subclasses within classied image data to nd connections between individuals, that share their models prediction and whether it can help to detect
outliers. Finally, it demonstrates, that a neural networks dimensionality could even be reduced through pruning, without losing overall performance based on the disclosure of inherent conceptual structure.