Complex sensor based systems are largely accessible by IoT based technologies and learn from user and machine generated data. These often called smart, or cyber-physical systems are well known and used in different everyday life settings. The majority of theses sensor based systems still rely on learning from pre-labeled data in a mostly supervised way. By doing so we neglect the fact, that unknown data dependencies and links can be present in the data, and changes of the data, or changes in the underlying system are likely to occur. In this workshop we focus on novel methods to enable embedded sensor- and actuator based systems to learn and detect, in a self determined and therefore unsupervised way, from autonomously collected data and dynamic changes in the data. In addition to core technical papers that deal with aspects like algorithms, approaches and software/hardware architectures, we also invite and welcome submissions that focus on the perception, interaction and trust of users in these systems.

In modern society, we are all surrounded by millions of embedded systems that collect, process, and communicate data about us. The pure amount of data that is captured by sensors makes it close to impossible to understand the data, nor to model the fundamental and highly interwoven knowledge and its relation that is present and captured in the data.

Abstraction, as one of the fundamental principles in computer science, will play a major role to understand, process and handle the data on a large scale. Not in all situations there is the need to have a comprehensive understanding of the data on the lowest level (i.e., fine grained activities of daily living, interaction of employees with machines, etc.). Impacting information is more on the level of (i) understanding changes in the data and (ii) detecting the appearance of novel artifacts. Future systems must have the ability to detect them and react accordingly to establish models for a lifelong learning and adaption process. Any static defined system that cannot cope autonomously with changes and drifts occurring in the applied application scenario will fail. Super- and semi-supervised methods that heavily rely on expert knowledge and training are the first step but must pave the way to systems capable of adapting and changing to occurring novelties to get around training data limitation and insufficiency.

AnNoSense asks questions on the potential and opportunities of Recognizing the Unknown – What’s behind the Black Hole of Big- and DeepData. To simply detect and predict outcomes is already helpful, but the challenge arises to explain, model, and describe the cause- and effect relation in both a theoretical and practical perspective. DeepData leads the path to combine data with the theory to understand and gain insights beyond what is already known. We hypothesise, that in contrast to state-of-the-art machine learning algorithms, as they search for the strongest signals in a pile of raw data and merely blindly encode the world according to statistics, rather than semantics, novelty and anomaly driven based adaptive models in combination with semantic graphs that react on novel occurring events and changes, are more suitable for a lifelong adaption and to understand the meaning of the underlaying signals. Identifying something as new or anomalous offers the possibility to draw conclusions and extend the experience that is captured in a model dynamically, and by far over the originally captured limited knowledge that is present in the initial training data.

The increasing complexity of machine learning models also pose challenges for understanding them, especially from the perspective of end users. The opacity of machine learning models has evoked new movements and research domains, which argue that AI decision systems must be able to transparently explain their behavior to all involved stakeholders. The European parliament states that AI systems should be “understandable to non-technical audiences [..] which is necessary to evaluate fairness and gain trust''. In this regard, new methods must be developed that to make complex sensor based “systems-of-systems” more transparent, but also include non-experts in the design and training process. This will be necessary to gain individual and societal trust and acceptance.