The DeepProbCEP system for Neuro-Symbolic Complex Event Recognition
KeywordsNeural-symbolic learning ; Complex event recognition ; DeepProbLog ; DeepProb- CEP ; Neural networks ; Symbolic inference ; Temporal reasoning ; Knowledge representation ; Deep learning ; MNIST
This research delves into the intersection of neural networks and symbolic reasoning, particu- larly focusing on the application of neural-symbolic learning and reasoning in Complex Event Recognition (CER). Central to this study is the exploration of DeepProbLog, a cutting-edge neural- symbolic framework. DeepProbLog distinguishes itself by adeptly integrating logical expressive- ness with the statistical strength of neural networks and probabilistic programming. While the Neuroplex framework also offers a robust blend of neural networks and logical programming, DeepProbLog was chosen for its enhanced ability to model uncertainty directly and to learn effi- ciently from sparse data. The strength of DeepProbLog lies in its ability to reason over high-level concepts and relationships using probabilistic logic programming, combined with its proficiency in handling sparse data through probabilistic reasoning. This framework enables the declarative definition of complex event patterns, facilitating nuanced CER by directly modeling and reasoning across diverse data representations. This thesis presents a thorough evaluation of DeepProbCEP, an extension of DeepProbLog, across various CER tasks in MNIST dataset. It includes a detailed comparative analysis against models rooted exclusively in either neural or symbolic approaches, highlighting their intrinsic strengths and limitations. The research offers valuable insights into potential future advancements in CER, focusing on the utilization of DeepProbCEP. While the primary focus of this research is on DeepProbCEP due to its capabilities in modeling uncertainty and learning from sparse data, the potential of Neuroplex within the CER domain is also acknowl- edged. This study contributes to the field by not only substantiating the promise of DeepProbCEP as a framework for CER but also by setting a foundation for future explorations and advancements in neural-symbolic learning and reasoning.