Multi-Task and Transfer Learning with Recurrent Neural Networks
Sigurd Spieckermann · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · 2015
The dynamics of complex technical systems can be approximated by recurrent neural networks (RNN). Such methods have proven to be powerful alternatives to analytical models which are not always available or may be inaccurate, but they often require large amounts of data, which is a scarce resource in many applications. In this thesis, RNN models are developed which allow for data-efficient knowledge transfer from source task(s) to a related target task that lacks data. The primary contribution is a novel RNN architecture which uses factored tensors to encode cross-task and task-specific information.