Deep learning models: duality, robustness and generalization properties

The ever-increasing scale of modern scientific and technological data sets raises urgent requirements for data-driven learning algorithms that not only maintain desirable prediction accuracy but also have high computational efficiency. However, a major challenge is that the data analysis and learning algorithms suitable for modest-size data sets often encounter difficulties or are even infeasible to tackle large-volume data sets. In this research large scale algorithms will be studied for supervised, unsupervised and semi-supervised data-driven modelling. A model class will be studied for which duality principles (e.g. in the sense of primal and dual model representations and kernel methods) are important. Scalability properties, stability of learning algorithms and generalization properties will be investigated.