A typicality based interpretation of neural networks: an experiment on facial emotion recognition

The thesis is developed around the application of a typicality-based approach to verify some properties of a neural net, in particular the multilayer perceptron. The subject we have chosen is the classification of basic emotions using the Facial Action Coding System (FACS) [3]. Action units are extracted from a dataset of labelled images using OpenFace [2]. The output, a vector of action units, is input to a multilayer perceptron for the classification problem. The network itself can be regarded as a weighted conditional knowledge base in a fuzzy typicality logic. The predictions of the net are used to construct a preferential model which is used to evaluate conditional formulas. We consider the finitely many-valued case and use Answer Set Programming (ASP) to evaluate typicality assertions to validate input/output properties of the net.