人工知能学会

Traditionally, when applying symbolic learning on low-level tasks such as pixel image processing, most systems rely on extracting knowledge from a preprocessing stage by using handcraft processors or other machine learning methods. Since logic programs outperform other machine learning models in generalizing from fewer data for their expressivity, it would be beneficial to consider using logic programs as a universal machine learning model to directly explain low-level tasks with rules. However, searching the space of logic programs efficiently remains a challenging problem. One approach is to split the problem into two interleaved parts: rule mining and parameter estimation of rule weights. Mapping logic programs into a continuous space allows global optimization, effectively pruning the huge combinatorial rule space. In this work, we propose an efficient algorithm for optimizing rule weights of programs with large amount of rules, showing the feasibility and potential of directly using logic programs as low-level machine learning models.