The Fraunhofer HHI researchers Sebastian Lapuschkin and Wojciech Samek received together with Grégoire Montavon and Klaus-Robert Müller from TU Berlin and Alexander Binder from SUTD the best paper prize for visualizing and understanding deep neural networks at the Workshop on Visualization for Deep Learning at the International Conference on Machine Learning (ICML) held in June 2016 in New York City, USA.
The award is endowed with a GeForce GTX Titan X from NVIDIA. The publication “Analyzing and Validating Neural Networks Predictions” summarizes the authors' work on Layer-wise Relevance Propagation, which is a newly developed technique to explain predictions of complex machine learning algorithms such as deep neural networks or kernel methods. The authors' work enables researchers to better understand what their highly non-linear models are actually doing and why they sometimes fail. This interpretability aspect is not only important in domains where relying on a black box algorithm is not acceptable (e.g., in medical applications), but it also may help to identify weaknesses of current state-of-the-art methods and consequently to improve them.
Sebastian Lapuschkin works for Fraunhofer HHI since September 2014. He is pursuing his PhD thesis on the topic of interpretability of machine learning algorithms. Wojciech Samek is head of the newly established research group "Machine Learning" at Fraunhofer HHI. The group works on different aspects of data analysis and has a strong focus on deep learning and interpretability.