A framework for improving offline learning models with online data

• Sabrina Luftensteiner
• Michael Zwick
• Georgios Chasparis
• Stefanie Brayer
• Thomas Grubinger

This paper proposes a framework for improving offline learning models with online data. The usage of online data is rising as machines get equipped with more sensors and therefore can produce more data, which can then be used directly for the fitting of already existing models to enhance predictions. In online learning, when fitting an existing model with new (online) data, catastrophic forgetting may occur. We propose a new framework that utilizes several state-of-the-art methods in deep-learning and machine-learning to minimize catastrophic forgetting through comparison. These methods range from memory-based approaches to methods for loss calculation and optimizers in deep learning. The proposed framework is specifically tailored for regression problems in the industrial field. It can cope with single and multi-task models, is easily expandable and enables a high variety of configuration possibilities for adaptation to the given problem.