Jean-Gabriel Ganascia - Research Activity
Since my appointment as full professor at the University Pierre et Marie Curie in November 1988 I have pursued my research on automatic machine learning in the LAFORIA research laboratory, which has since become part of the Computer Science Research Laboratory of Paris 6 University (LIP6). I have gradually put together a research team, ACASA, whose name means 'house' in Portuguese and in fact stands for Acquisition des Connaissances et Apprentissage Symbolique Automatique (Knowledge Acquisition and Automatic Machine Learning). The activities of the group follow implicitly from the name, ie we try to use automatic machine learning techniques to facilitate knowledge acquisition. However, taken literally the research activities might seem confusing, since in every day language all learning is a question of knowledge acquisition and vice versa. In order to understand exactly what my research team is doing, it is therefore necessary to explain what is meant, in the technical sense, by "machine learning" and "knowledge acquisition". What is more, our research subjects have evolved and now focus essentially on scientific discovery, on discovery in data bases and even on creativity.
Machine Learning and Knowledge Acquisition
Automatic machine learning includes a number of algorithmic techniques the aim of which is to enable machines to learn, either in order for the machines to use their experience so as to execute routine tasks more easily or so that they can build the knowledge bases of expert systems or so that they can retrieve knowledge from any knowledge bases. As for knowledge acquisition, it first appeared as a form of knowledge-based system engineering, in other words as a set of methods used to build expert systems. Then, gradually, it began to concern knowledge modeling and now designates both the acquisition of knowledge by humans using machines and the acquisition of knowledge for machines, using the knowledge and know-how of human experts of the various domains. In all these cases machine learning techniques are extremely useful for knowledge acquisition and this is what underpins the scientific research of the ACASA team I head at the University Pierre et Marie Curie (Paris VI) in the computer science laboratory (LIP6).
Contribution to Machine Learning
In the field of machine learning, at the end of 1992 I made available to my research team the software application CHARADE that had resulted from my PhD on association rules. Between 1993 and 1996 I had students test the software on large size applications. Isabelle Moulinier and Gailius Raskinis tested it on text categorization; three others tested it on data mining, also called knowledge discovery in data bases: Alfred Attipoe worked on data mining in the medical domain during his postgraduate degree work placement, Vincent Corruble used it during his doctoral research, and Anca Banto tested it during her postgraduate degree work placement.
With Jean-Daniel Zucker I extended the functions of CHARADE by coupling it with a particularly efficient structural matching mechanism that used the original notion of morion.
Lastly I generalized the underlying theory which relied initially on a Galois match between boolean lattices so that it would rely on a Galois match between distributed lattices. Thanks to this generalization it was possible to understand the privilege given to the notion of attribute in the learning systems written in propositional logic. (Download the articles published in the proceedings of the IJCAI93 and ALT93 conferences)
Having familiarized myself with the use of lattices in learning, by 1988 I was designing original conceptual clustering techniques on Sowa graphs. Isabelle Bournaud started to program and test these techniques in 1993 and in 1996 successfully completed and defended her PhD.
More recently, I worked on conceptual clustering from texts with Julien Velcin.
Contribution to Knowledge Engineering
In the field of knowledge acquisition, my contribution from 1993 to 1996 was above all as a sub-contractor as part of the ESPRIT VITAL contract, and it was ONERA (Office National d'Etudes et de Recherches en Aéronautique) who funded two PhD students, Jérôme Thomas and Bernard Leroux, whose research I supervised. (Download the ECML'93 article in postscript)
We developed a symbiosis between the knowledge acquisition models inspired by the KADS methodology and the input of the CHARADE system, for which these learning models present a clear formulation of learning bias. The result was a software program developed by Jérôme Thomas, who completed a doctoral thesis on the same subject.
Other work was carried out in collaboration with ONERA and the Bank of France and examined more deeply the notions of knowledge model as promoted by knowledge acquisition during the eighties. This research, which was carried out in collaboration with partners involved in practical applications, gave rise to two doctoral theses, those of Bernard Leroux and Catherine Vicat. (Download the KAW'95 article)
The work of Isabelle Bournaud, which has already been mentioned and which dealt with the clustering of conceptual objects described by Sowa graphs, was directly related to the work done with Esma Aïmeur and then Catherine Faron on taxinomic knowledge acquisition expressed in the form of Sowa graphs. It is the very accumulation of large quantities of structured objects which provides the justification for the attempts at conceptual clustering on such objects. (Download the ICCS'95 article)
Approaches to scientific discovery and creativity
The final and surely the most original aspect of the research carried out by my team during the last four years is the acquisition of knowledge by human with the help of machines.
We studied scientific discovery in medicine with Vincent Corruble, the aim being to understand the status of induction in medical knowledge before the 19th century. This research was done thanks to computer models built using the machine learning tools that had been developed by our research team. An article describing the research was published in the Artificial Intelligence Journal.
We also began to explore the use of machine learning techniques developed by our team in order to produce computer-assisted discovery in the domain of medicine, using existing data bases.
Similarly we explored computer-assisted discovery in the domain of molecular biology, with Irina Tchoumatchenko, and in the domain of music for the detection of recurrent themes in the music of Charlie Parker with Pierre-Yves Rolland, who is currently doing a PhD under my supervision.
Lastly, if it is true to say that certain artistic activities such as musical accompaniment require certain skills, then these activities can be modeled using knowledge engineering techniques, and then simulated on a computer. Daring indeed, yet it is just pushing to their logical conclusion the ambitions of knowledge engineering specialists who claim to be able to address all professional knowledge and skills, first to model them and then to simulate them on a computer. I developed a model of creativity based on the notion of memory and broadly inspired by the work of Robert Schank. (Download the commentaries on the Schank model) I then applied this model to the simulation of the improvisations of a double bass player of a rhythm section of a jazz band, with a student Geber Ramalho, whosuccess fully defended his doctoral thesis on the subject in January 1997.
In addition to purely scientific research, I have also in the last few years explored the question of philosophical approaches to induction. After all, machine learning is, to a large extent, inductive, ie it moves from the particular to the general. If we look at the history of philosophy, and more precisely at Aristotle and other equally prestigious philosophers including Bacon, Leibniz, Buffon, Reid, Mill, Peirce, we find a strong tradition of the study of induction. And while laying down the philosophical status of inductive reasoning, this tradition also attempted to determine the fundamental logical mechanisms involved. Consequently, there is a meeting point between machine learning, which tries to simulate inductive reasoning using formal manipulations, and the tradition of the philosophical study of induction. We have tried to identify this meeting point and to show that contemporary philosophical thinking about induction, with the limits of inductive reasoning, do not impact on the prospects of artificial intelligence that have been opened by machine learning.