Modern applications, such as maritime situational awareness, require the processing of large data streams that are continuously evolving over time. Complex event recognition (CER) systems reason over high-velocity data streams containing `simple, derived events’ (SDE)s, i.e., symbolic abstractions of raw sensor data, in order to detect instances of `complex events’ (CE)s, i.e., spatio-temporal combinations of SDEs, by means of matching a set of pre-defined patterns over incoming SDEs. In some applications, the input stream might consist of sub-symbolic data such as images or SDEs associated with some uncertainty.
Recent advances in neurosymbolic AI aim to bridge the gap between symbolic AI, which provides interpretable reasoning and formal representations, and sub-symbolic approaches such as deep learning, which excel at perception and pattern extraction from raw or uncertain data. By integrating these paradigms, neurosymbolic systems can enhance trustworthiness, transparency, and explainability, properties required in safety-critical domains where human operators must understand and validate system decisions.
For this thesis, the candidate will explore and/or develop a wide range of techniques for performing complex event recognition on uncertain or sub-symbolic data streams. The work will emphasize on combining deep learning and symbolic reasoning approaches to develop a neurosymbolic system for complex event recognition that is capable of providing accurate, and interpretable CE detections. The developed system may be evaluated on a variety of application domains, such Maritime Situational Awareness and Human Activity Recognition.
Bibiliography:
https://cer.iit.demokritos.gr/publications/papers/2015/artikis-TKDE14.pdf
https://www.sciencedirect.com/science/article/pii/S0957417422023946