Abstract

Deep learning is a particular kind of machine learning which is powerful and
flexible as a consequence of its ability to represent the world as a nested hierarchy
of concepts (LeCun et al. 2015, Goodfellow et al. 2016). Due to the ever increasing
power of parallel computing graphics processing units, larger labelled datasets and
improved training techniques, great leaps in the performance of various machine
learning tasks have been achieved using deep learning (LeCun et al. 2015). At the
time of this writing, deep learning is the dominant machine learning approach for
much ongoing work in fields such as: Computer Vision (Krizhevsky et al. 2012, He
et al. 2016), Reinforcement Learning (Mnih et al. 2015, Silver et al. 2016), Medical
Imaging (Ronneberger et al. 2015), and Natural Language Processing (Vaswani
et al. 2017, Devlin et al. 2019).
However, with the uptake of deep learning models into safety-critical domains,
transparency of model predictions is becoming increasingly important for: safety,
decision-making, fairness and legislative reasons. Moreover, designing deep learning models that strike a good balance between human interpretability and performance has proven to be a challenging task (Caruana et al. 2015, Montavon,
Lapuschkin, Binder, Samek & Müller 2017, Kendall & Gal 2017, Rudin 2019,
Samek et al. 2019). With that said, in this thesis we advocate for an alternative
view of interpretability based on estimating the uncertainty in a model’s predictions, which serves as a proxy for model transparency. In our investigations,
we formalise the desiderata of model transparency as: trust, information and
generalisation, and take steps towards the development of deep learning models
which have the potential to satisfy them. Concretely, we leverage the language of
uncertainty to improve the performance and transparency of deep learning models in computer vision tasks, providing probabilistic techniques to enhance more
interpretable models by design such as capsule networks.