Logo der PTB - Messen, Forschen, Wissen

Deep learning

Working Group 8.42

Overview

Deep learning belongs to the class of machine learning methods and typically employs neural networks with many layers to solve tasks such as classification or function approximation. Due to their flexibility neural networks are widely applicable and have shown extraordinary performance, for example in autonomous driving, computer-aided diagnosis, or the automatic segmentation of images. To make deep learning applicable for metrology, it is crucial to understand and evaluate the reliability of these methods. One important issue in this regard is to quantify the uncertainties associated with their results. Models employed in metrology are usually well understood and often based on physical knowledge. Deep learning, on the other hand, builds its models directly from data. Another challenge in the application of deep learning for metrology therefore is to understand the behavior of the empirically built models.

Example for regression by a neural network and quantification of the associated uncertainty.

Research

One focus of PTB’s Working Group 8.42 is on methods for explainability and uncertainty evaluation of results obtained by deep neural networks. A current development is the use of the Fisher information to inspect adversarial examples. Understanding the behavior for these specifically designed inputs is of fundamental importance to judge and enhance the reliability of neural networks. Another focus is to employ deep learning for applications within metrology. Examples include the assessment of image quality in X-ray diagnostics or computational optical form measurements.

Software

Publikationen

L. Hoffmann, I. Fortmeier;C. Elster
Machine Learning: Science and Technology,
2021.
J. Martin;C. Elster
Appl Intell,
2020.
T. Kretz
PhD Thesis
, 2020
L. Hoffmann;C. Elster
Journal of Sensors and Sensor Systems, 9
301--307,
2020.
T. Kretz, K.-R. Müller, T. Schäffter;C. Elster
IEEE Transactions on Biomedical Engineering,
2020.
J. Martin;C. Elster
Neurocomputing, 382
80--86,
2020.
Export as:
BibTeX, XML

Preprints