Continuous learning on multimodal data streams

Intelligent systems must master robust yet flexible learning strategies for everyday use: New, unforeseen situations constantly arise for which data from different sources must be merged and evaluated in a short time. The usual heterogeneity of data poses a particular challenge. For 3D scenes, for example, image data, depth data and textual descriptions are available, while in medical diagnostics laboratory findings, measurement series and imaging procedures have to be dealt with simultaneously. The aim of the research project is to develop learning procedures for multimodal data with which the knowledge base can be continuously adapted across domains (transfer learning) and applied to new content without jeopardising the quality of the inference (lifelong machine learning).