Drift-Learning for Diagnostics

Anton Pfeifer received his Master’s degree in Information Technology from the Technische Hochschule Ostwestfalen-Lippe in Germany. Currently he is working towards his doctoral degree as research assistant at the inIT - Institute Industrial IT in cooperation with the Brandenburg University of Technology Cottbus-Senftenberg. His main research topics include machine learning, information fusion systems, and signal processing in the context of mobile health systems.

Anton Pfeifer

Machine learning has recently grown in popularity and importance. Especially areas of pattern recognition and natural language processing have been revolutionised by Deep Learning. Deep Learning describes algorithms based on artificial neural networks that learn from experience. The algorithms are applied today in everyday life for image and speech recognition or in industry for automated defect and anomaly detection. Most algorithms are based on a gradient-based approach called backpropagation. Its current form was first published in 1970 by Seppo Linnainmaa and was first applied to neural networks in 1982 by Paul Werbos. Today it is applied to successively weaken specific connections of neural networks and to strengthen others. As a result, neural networks learn progressively from their own errors.

Relevant research progress has been made on modelling relations between observations and latent representations of them by Huang and LeCun in 2006. These so called convolutional neural networks (CNN) have been proposed for considering spatial dependencies, such as shapes and edges of an image. In the analysis of real-valued signals (for example, images or audio signals) the data are often multidimensional. Therefore, it requires a specific approach to consider relations within the data. Parcollet et al. showed in 2019, for example, that a real CNN cannot extract the colour information when trained on grayscale images. This makes it unsuitable under heterogeneous conditions. Therefore, neural networks based on multidimensional numbers were proposed. One of these approaches is related to complex-valued neural networks, where the magnitude and the phase are embedded in a complex number. The specific complex algebra allows complex-valued neural networks to preserve the relation between the magnitude and the phase during the learning.

Recent advantages on the field of multidimensional neural networks make it possible to build state-of-the-art architectures, such as complex CNN, to efficiently solve tasks such as image processing. However, the field of multidimensional neural networks is still a resurgent field. Therefore, this work will investigate three promising directions: (i) It is necessary to evaluate new data pre-processing methods to transfer the characteristics of real-valued data into the complex domain. Thus, the physical nature of the signal is completely preserved. (ii) Although new complex convolutional neural networks are proposed, novel architectures are still missing. For example, capsule networks or generative adversarial networks. (iii) An implementation in current frameworks could drastically reduce the computing time and make multidimensional neural networks an alternative to real-valued neural networks.