Fraunhofer Center for Assistive Information and Communication Solutions – AICOSThe Intelligent Systems (IS) group is formed by a dynamic, multidisciplinary, and innovative research team specialised in developing intelligent ICT solutions that leverage AI, computer vision, time series, and natural language processing to address real-world challenges. The team strives to push the boundaries of AI and related fields, collaborating closely with industry partners, academia, and other research institutes to foster innovation and create practical applications and solutions that positively impact society, industries, and individuals. With a main focus on health and production industry, we support our clients throughout the entire research and development process. From defining a suitable approach and roadmap to address their specific business problems to collecting, analysing, refining, or even synthetically generating data, we ensure a solid foundation for our AI-driven solutions. Moreover, we are committed to providing reasoned, ethically responsible, and trustworthy AI decisions, recognising the importance of transparency in implementing AI technologies. Finally, we support the seamless integration, monitoring, and maintenance of intelligent solutions, enabling our clients to harness the full potential of AI.
Data-driven intelligence
How can we extract meaningful insights from high-dimensional, unstructured data? This sub-field aims to address this question by leveraging our strong background in classical time series and image processing within machine learning. We focus on char-acterizing unstructured data, identifying patterns, predicting events or longitudinal trends in time series, and extracting relevant image features, including objects, sub-structures, patterns, and textures. Furthermore, we combine different modalities to obtain a broader understanding of the phenomena under study. Our research also encompasses anomalies, out-of-distribution detection, and domain-driven data quality evaluation in real-world and synthetic scenarios. We articulate with HCD in mapping and utilizing domain knowledge and with CT in data engineering developments.
Learning in low-data scenarios
How can we address the real-world problem of low data availability? This sub-area focuses on developing advanced learning techniques that can operate effectively in low-data scenarios. We explore few-shot learning, transfer learning, and fine-tuning of large models to maximize the utility of available data. Additionally, we investigate generative models for data synthesis, creating new data that can help improve model performance. Our research also extends to distributed and federated learning, which allows us to leverage data from multiple sources while preserving privacy and reducing data movement.
Trustworthy and Operational AI
How can we ensure that AI models are accurate, reliable, interpretable, and easily deployable in real-world applications? This research line is dedicated to developing robust AI pipelines that can withstand data perturbations and distribution shifts. We focus on privacy-preservation techniques to protect sensitive information, and uncertainty quantification to provide a measure of confidence in our models' predictions. Our work, in tandem with HCD, also includes human-centred and explainable AI, by meaning-making and ensuring that our models’ decisions can be understood by humans. Lastly, together with CT, we aim to develop models that can be deployed on the edge, bringing the power of AI closer to where the data is generated and used.