Lifelong machine learning
Contact person: Prof. Dr. Lambert Schomaker, Dr. Hamidreza Kasaei
Current examples of deep learning application are often impressive, but based on closed data sets. Such data is very clean, heavily curated in terms of segmentation (i.e., the selection of relevant material) and labeling. This is in stark contrast with real-life problems in AI-based robotics and in industrial processes. Here, the amount of labeled material is usually limited and class labels are varying over time. There may be drifts in raw-data properties (sensors, pre-processing methods that are varying over time, etc.). Such problems are not solved, at all. Application domains are in historical document analysis and robotics.
|Last modified:||17 November 2021 3.53 p.m.|