< Back to previous page

Project

Combined Reasoning and Learning Approaches

Goals. To address the limitations of the existing approaches that lie in the intersection of ML and KRR as outlined above, the goal of this thesis is three-fold. First, we aim at developing advanced scalable algorithms that allow for incorporation of rich background knowledge into ML systems represented as Description Logic ontologies or Answer Set Programs. Second, we seek for exploitation of possible approaches for bringing the rule induction methods to the industrial scale. Last but not least, the goal of this thesis is to adapt the methods developed within this thesis for various applications within the field of manufacturing (e.g., scheduling, repairing failures in production pipelines where ML methods could be combined, e.g., with [5]) and beyond (e.g., autonomous driving). Potential approaches. The first research question to be exploited within this thesis concerns the tight integration of deductive reasoning methods from KRR into ML models. We start with exploring this topic in the context of Knowledge Graph embeddings and further extend it to other settings. KG embedding models recently demonstrated very promising results on learning a 'good' representation of KGs by embedding vectors of entities and relations. They can be effectively used for fact prediction and answering conjunctive queries over incomplete KGs [23]. However, these methods are still limited in capturing rich background knowledge. To address this shortcoming, we plan to use symbolic reasoning to 'guide' the embedding models to improve themselves and to mimic/learn how to perform reasoning over KGs efficiently. As a starting point we plan to focus on KRR formalisms with good computational properties for which effective reasoning methods exist [4]. One can either use KRR techniques for regularizing the ML models or to integrate their results directly into the optimization process during the model training. Moreover, the findings from [1] could be helpful to better understand the differences between the symbolic and embedding approaches towards their effective combination. The second research direction to explore is related to the optimization of methods that present a loose coupling of KRR and ML techniques. In this context we plan to look into the combination of Answer Set Programs and Deep Learning methods [17] and aim at improving the efficiency of their interaction. One of the important issues in this context is the grounding process that the ASP systems have to perform. This is the main computational bottleneck of the current methods. To address the grounding issue we propose to interlink the grounding process with solving and perform the interaction with ML models in an online fashion rather than sequentially. Another research direction that nicely complements the above mentioned ideas concerns the automatic acquisition of rules from the data. Indeed, while formal representation of background knowledge can be manually specified, this process is often tedious and thus calls for automation. For example, extraction of temporal rules from the manufacturing data is particularly appealing in the context of production pipelines. Learning such rules as 'The engine cannot start before it is switched on' could be helpful for verifying the correctness of manufacturing processes. We also see a high potential in learning dispatching rules in the scheduling domain as well as learning rules from the input-output of Deep Learning systems to contribute to their interpretability.

Date:25 Mar 2021 →  Today
Keywords:Knowledge graphs and representation, neural networks
Disciplines:Knowledge representation and machine learning
Project type:PhD project