Jensen Huang, CEO of NVIDIA, the world leader in computer graphics and AI computing, recognized 12 internationally outstanding works in the field of machine learning at the Computer Vision and Pattern Recognition Konferenz (CVPR) in Salt Lake City.
Explainable AI - Understanding Neural Networks
The novel analysis method developed by the DFKI researchers allows far-reaching insights into the properties and processing procedures of deep neural networks that are otherwise largely opaque. The processes are observed from the outside by a second neural network. A powerful autoencoder receives feedback on the results of the analysis and adjusts the input accordingly, thus reducing errors. For example, if the network does not correctly recognize a motif in an image, the input is adjusted according to the network's taste. From this, conclusions can be drawn about the preferred decision paths of the most common models and these can be visualized in an understandable way. In short, the procedure brings a new light into the processes of the "black box" of deep learning and thus helps to make the processes more comprehensible. This is a central aspect if machines are to make decisions in the interests of people.
In addition to the DFKI, the only prizewinner from Europe, other renowned institutions such as Stanford University, Tsinghua University, University of Toronto, University of Tokyo and University of Washington, Chinese Academy of Science, Peking University, were also honored.
Prof. Dr. Andreas Dengel, site head and Director of the Research Department Smart Data & Knowledge Services at DFKI in Kaiserslautern, is delighted about the renewed award: "We are very proud that our research work has once again been recognized by the world market leader for machine learning platforms. In particular, I would like to congratulate the employees of the DFKI Competence Center Deep Learning who are regularly awarded for their excellent work".
Sebastian Palacio, scientist in the DFKI Smart Data & Knowledge Services team, presented the new concept for the first time at the CVPR conference with the paper "What do Deep Networks Like to See?" (DFKI authors: Sebastian Palacio, Joachim Folz, Joern Hees, Federico Raue, Damian Borth and Andreas Dengel) and accepted the award.
The technology was developed within the DeFuseNN project funded by the Federal Ministry of Education and Research (BMBF).
Further Informationen:
Link to the paper
