Here is a follow-up of my previous summary about the 25th Conference on Inductive Logic Programming that I have just attended in Kyoto. The current post is not self-consistent, in the sense that it should be read alongside with my summary of day 1 of ILP’2015 (to illustrate it, the numbering of this post is a continuation of the former one).
I will focus here on day 2 and day 3, to conclude with general considerations about my feelings about this community and the organization of the conference. As a reminder for those willing to know more, organizers uploaded all papers on ILP website.
6/ Morning of day 2 focused on probabilistic approaches in learning. Taisuke Sato, who gave a talk about the distribution semantics he proposed 20 years ago, insisted on his conviction that probabilistic logic learning and statistical relational learning will play a major role for machine learning techniques to be able to tackle huge and noisy data, and he showed how his distribution semantics is a way to address such issues.
7/ The third – and last – invited talk, in day 3, from Luc de Raedt, also discussed probabilistic approach, more specifically probabilistic logic programming. From ProbLog (which is a probabilistic Prolog), various applications were given, from the identification of the magnetic (or not) nature of objects based on observation to biological analysis of the connection between genes in a graph.
8/ A significant number of talks emphasized the need to benefit from modern computational structure (cloud computing, clusters, …) to make the ILP approaches be able to address large-scale data. As a consequence, it is necessary to develop specific algorithms that are able to take profit from the new available infrastructures. Different methods are proposed, like the use of a MapReduce approach, specific contributions to use GPU kernels or alternative parallel-processing systems.
9/ In some applications (e.g., Luc de Raedt presented a nice and challenging case study for a robot to recognize graspable points), it is crucial to process relational kernels on graphs. Approaches around kernels arouse much efforts from the community.
10/ I have been struck by the wide range of applications of ILP approaches. While ILP, and more generally logics, is still considered as a difficult topic that many bachelor or master students fear, it definitely has many practical, useful and elegant applications. Obviously systems biology, in which I have a special interest, but also robotics (e.g., rescue robots), medicine (to analyze the correlations between medical records and illnesses), object recognization, … All these topics are really exciting, as they question our capacity to abstract the world in terms of facts, rules and more generally logics.
11/ This session of ILP ended on a panel celebrating the 25th anniversary of the conference. The chairs of the five last editions shared their views about the current trends and next outcomes for ILP. Many topics and keywords about it were discussed. Among them are the emergence of new application fields (Stephen Muggleton for example quoted synthetic ecology), the huge challenge of deep learning, or the improvement of parallel-techniques to guarantee ILP techniques can process real-life applications on a daily basis.
12/ As a first-time attendent to ILP, I was deeply interested in the quality of the talks. The mix between conceptual researches and practical illustration of the merits of such works makes the conference to be accessible.
13/ As a final word, the conference’s lunches and social dinner has contributed to make it memorable. Tasty bento, and a dinner that allowed everyone to experience being served by maikos and listen to classical shamisen dances and songs. This really was what the Japanese call omotenashi (おもてなし), i.e., the sense of hospitality.
Morgan Magnin 