Improving generalization of neural networks via unsupervised domain adaptation and semi-supervision

Advances in machine learning (ML) have accelerated research in biomedical
image analysis and promise to disrupt clinical practice by enabling tools for
improved diagnosis and treatment pipelines. However, the performance of ML
systems often degrades when they are applied on test data that differ from the
training data, for example, due to variations in imaging quality and protocols.
This hinders usage of ML systems on large-scale studies or outside the lab.
The first part of this talk will introduce the issue of domain shift between
training and testing data distributions and the basic theory of domain
adaptation that tries to alleviate it. It will then cover our work with adversarial
networks for unsupervised domain adaptation, which investigated their
potential for segmentation of lesions in multimodal MR images and proposed
their extension with multi-connected networks for improved adaptation. It will
then provide insights in the behaviour and limitations of the approach.
Motivated by a common limitation of unsupervised domain adaptation
methods, that they do not account for discriminative separation of the
unlabelled target samples, the second part of the talk will focus on the
complimentary problem of semi-supervision for a solution. After discussing the
low-density separation theorem I will present our recent work on regularization
of neural networks by capturing the manifold of unlabelled data in latent space
via graphs and enforcing a network to learn more discriminative
representations.