doc. Boris Flach, Dr. rer. nat. habil.

The application of deep learning in computer vision domains such as earth observation or medical imaging is often hampered by the lack of annotated training data and requires semi-supervised learning. One possible route to novel semi-supervised learning approaches in deep networks is to provide them with feedback connections. This is also supported by recent neuro-cognitive evidence for the importance of feedback connections in visual processing. The aim of this research is to develop novel semi-supervised learning methods for deep networks with complex generative feedback. This will require methods beyond stochastic gradient descent for standard error backpropagation. Research should focus on novel semi-supervised learning methods with provable convergence and approximation guarantees. The development will start from statistical machine learning and probabilistic neural networks. It will include advanced methods from convex & non-convex optimisation such as primal-dual optimisation, block coordinate descent, multi-criteria optimisation and others.