Invertible Neural Networks with Applications in Computer Vision

The focus of the PhD project is on the development and refinement of invertible neural networks with applications in the field of computer vision. Neural networks, and convolutional neural networks in particular, have proven to be an extremely powerful tool in computer vision among many other large-scale applications. Some computer vision tasks are highly relevant to the modern insurance business, for example: (1) using damaged car images from the customers for damage detection and identification in order to expedite the motor claim process; (2) using indoor and outdoor property images to learn useful latent features to improve the modelling of home insurance pricing.

Invertible Neural Networks (INNs), as compared to classical neural networks, enjoy several additional advantages. Firstly, they have the potential to achieve accuracies for large-scale applications that go well beyond standard networks. This is due to their strongly favourable memory footprint, which allows for the training of larger (deeper) networks and higher-dimensional input data as compared to standard deep learning settings. As such, INNs would enable the insurance business to scale to larger data sets hence wider applications in the global insurance business. Secondly, INNs allow for training so-called 'normalising flows' - powerful models for learning complex probability distributions from samples, which allow for both likelihood estimation as well as efficient sampling. The former is particularly applicable to risk assessment for insurance policies.

While INNs are readily applicable to many applications relevant to the insurance business, they suffer from a lack of guarantees in terms of regularity and numerical stability. Without these properties, one cannot ensure that the models are trustworthy thus reliable for large-scale applications. The project thus aims to develop the regularity and stability theory of INNs further, to allow for more fine-grained stability guarantees.