Presenter

Mr Meshia Cedric Oveneke - ETRO, Vrije Universiteit Brussel [Email]

Abstract

Empathy is one of the most fundamental cognitive building blocks of human intelligence. What if empathy was as fundamental for artificial intelligence? Motivated by this conjecture and the fact that artificial intelligence systems will continue to play a bigger role in all aspects of our society, the present dissertation aims at contributing towards computational empathy, with a focus on its core cognitive component: the capacity to continuously understand another’s affective state from facial displays thereof. Continuous affect estimation from facial expressions has attracted increased attention in the computer vision and machine learning research communities. At its heart are a set of fundamental information processing challenges caused by the complex deformation of the face, the presence of nuisance factors (e.g. scale, orientation, lighting conditions), and the lack of reliable ground-truth data.

The main objective of this dissertation is to design a principled framework for automatically estimating continuous affect from video sequences. To this end, two sets of research questions are formulated: (i) computational theory, i.e. what is the goal of the computation, why is it appropriate and what is the strategy by which it can be carried out? This set of questions is addressed by leveraging the Bayesian filtering paradigm, i.e. considering affect as latent dynamical system corresponding to a general feeling of pleasure with a degree of arousal, and recursively estimating its state using a sequence of (noisy) observations. (ii) representation and algorithm, i.e. what is the representation for the input and output, and what is the algorithm for the transformation? This set of questions is mainly addressed by leveraging the much-celebrated deep learning paradigm, i.e. automatically learning to extract meaningful representations in a deep and hierarchical manner.

From a technical point of view, the dissertation advances the state-of-the-art as follows: (1) canonical face representation (CFR): a set of novel algorithms for monocular 2D face frontalization, 3D facial shape-from-shading and scene flow estimation. (2) convex unsupervised representation learning (CURL): a novel frequency-domain convex optimization algorithm for unsupervised training of deep convolutional neural networks (CNN)s. (3) deep Kalman filtering (DKF): a Kalman filtering-based algorithm for affect estimation from a sequence of deep CNN observations. The performance of the resulting CFR-CURL-DKF algorithmic framework is empirically validated by means of qualitative and quantitative experimental results on publicly available benchmark datasets for facial expression recognition and continuous affect estimation. We believe that the outcome of the present dissertation will pave the way for contributing towards computational empathy and next-generation artificial intelligence systems.

Short CV

MSc. Applied Sciences and Engineering: Computer Science (Artificial Intelligence), 2013, Vrije Universiteit Brussel