< Back to previous page

Project

Adaptive and continual learning for context drift detection and mitigation in industrial applications

Traditional machine learning algorithms typically consider static environments, in which the data is drawn i.i.d. from a stationary probability distribution. However, the assumption of stationarity does not hold in most real-world applications, in which the data distribution changes over time due to the dynamicity of the environment. The state of the data distribution is determined by the context of the environment, due to which changes in the distribution are referred to as context drift. This drift can influence the distribution of the input variables, the distribution of the output or both. The latter two cases are instantiations of concept drift. In order to guarantee robustness against context drift, adaptive learning is able to perform (near-) real time model updates. In that sense, adaptive learning algorithms can be seen as advanced incremental learning algorithms that are able to adapt to evolution of the data distribution over time and update the resulting models online during their operation to react to change. More recently, continual learning aims to extend this principle by accumulating the learned knowledge over time, coping with the interference between the new knowledge and what was previous learned, without the need to retrain from scratch. To operate in such dynamic environments, models are required to detect context drift as soon as possible, while distinguishing it from noise, and if necessary, adapt as swiftly as possible. This robustness to change is increasingly challenging in industrial environments, in which machines and processes operate in inherently dynamic environments and moreover are influenced by a variety of different contexts. Therefore, this thesis will research novel adaptive and continual learning methods for context drift detection and mitigation in industrial applications. Particular attention will be paid to the interpretability of the approaches, in order to provide the domain expert insights about the influencing contextual factors triggering model updates, as well as their efficiency, since the methods often need to operate in resource-constrained environments.

Date:25 Nov 2021 →  Today
Keywords:change detection, time series analysis, adaptive learning, continual learning, concept drift, machine learning
Disciplines:Machine learning and decision making
Project type:PhD project