Identifying Influential Inputs in Probabilistic Logic Programming
Combining probabilistic theory and statistic reasoning, probabilistic logic pro- gramming (PLP) has various applications on many data-driven problems. Notable examples include machine learning models such as probabilistic soft logic (PSL) and Markov Logic Network (MLN). Despite the prevalence of PLP, there are few tools for us to debug and analyze PLP programs. In our prior work, we developed a PLP program debugging system which we call P3 for ProbLog. In this thesis, we further extend the system to more general PLP programs where the rule weights represent not only probabilities but also parameters in Markov Random Fields. Our PLP analyzing systems are based on probabilistic provenance which records the derivations of queried tuples. Based on the probabilistic provenance, we provide 3 types of queries: expla- nation query for inference derivation, probability query for calculating conditional probabilities, and influence query for identifying influential inputs. We evaluate our system on several PLP cases, such as Smoke and Visual Question Answering (VQA), and the evaluation results demonstrate the effectiveness of our system.
MetadataShow full item record
Showing items related by title, author, creator and subject.
Barsaloux, Kevin C (Georgetown University, 2017)Current evaluations of career technical education (CTE) programs seeking to assess labor outcomes for individuals are missing a generalizable model of program quality. Studies either focus on differences between CTE ...