Unsupervised Concepts Extraction in Neural Networks

Abstract

Recently, the domain of Explainable Artificial Intelligence (XAI) saw the advent of Testing with CAV (TCAV). Although very practical as they allow to check the importance of a concept in the decision making of a neural network, they pose the prerequisite of knowing the concept at stake and having a dedicated dataset.

In order to remedy this, a method is proposed to obtain an overview of the different concepts that are important in the decision making process of a neural network.

The idea beneath the proposed method is to compare and label the neural network instances based on their activation at a given layer inside the network itself.

Using two different kinds of neural network, an image classifier and a game agent, the method is tested to see if an unsupervised extraction of concepts is possible.

Date of Award	24 Jun 2021
Original language	English
Awarding Institution	University of Namur
Supervisor	Benoît Frénay (Supervisor)