B1633
Title: Assessment and calculation of model complexity of deep neural networks
Authors: Rene-Marcel Kruse - University of Goettingen (Germany) [presenting]
Benjamin Saefken - Clausthal University of Technology (Germany)
Thomas Kneib - University of Goettingen (Germany)
Abstract: Model selection is an area of research in the field of statistics that receives a lot of attention, with the concept of model averaging being revisited frequently. The idea of qualified and quantifiable model selection, however, so far has received little attention in the field of learning-driven methods such as machine and deep learning. We focus our attention on bringing concepts such as model complexity and degrees of freedom to the field of deep learning to derive a means to leverage the insights to perform data-driven model assessment and model averaging of deep learning models. We illustrate the theoretical and practical problems of translating these statistical techniques to the domain of deep learning. Further, we apply the proposed methods to examples of simulation studies as well as applications based on real-world data to illustrate the validity of our approach.