On Machine Learning, The Mathematical Toolbox Behind Data Analysis

Cuturi Marco

Cuturi Marco准教授My name is Marco Cuturi, and I have been appointed as Associate Professor in the Intelligence Science and Te c h n o lo g y De p a r tme n t of the Graduate School of Informatics on November 1st 2013. Please forgive me for using english in this text: although I have no trouble communicating in japanese with my colleagues for work related matters, I am not confident enough in my writing ability to write this self-introduction in japanese.

My work as a researcher in Kyoto University is in the field of machine learning. Machine learning has recently received much attention in the general media, because of its important role as the analytical engine of the "big data" revolution. Machine learning is a pluridisciplinary field which is difficult to define, since it blends various elements of computer science, statistics, mathematics (functional analysis and optimization) and neuroscience.

The goal of machine learning is, however, relatively easy to explain: because of the ever increasing presence of digital devices (smartphones, wearable computers soon) in our daily interactions (shopping, transportation, entertainment, work) and because of the partial digitalization of many aspects of our life (health, social interactions), vast amounts of data related to our daily life are now curated in huge databases. The goal of machine learning research is to leverage these databases to learn new insights (applicable to science, commerce, health, or more generally society) and create algorithms that will be able to manage that data, sometimes in real time, to produce relevant outputs automatically, with minimal maintenance. In the landscape of "big data", machine learning only plays a role at the very end of the information chain. Rather than studying how to collect and/or efficiently organize huge amounts of data, which are very important endeavors by themselves, I am only interested in the analysis of data itself.

Machine learning offers a contrast between, on the one hand, its claim to be built on clean mathematical foundations and sometimes contribute to that field, and, on the other hand, its aim to be directly useful for society. This contrast is the main driver of research in machine learning: researchers in my field must be willing to update their core scientific beliefs several times along their career, depending on what is the most pressing issue or the most promising approach at any point in time. This may for instance depend on progress in hardware (recently Graphical Processing Units), in mathematics (optimization) or the emergence of new problems (ranking answers for a search engine, predicting labels in large graphs).

For instance, some of the older and most successful researchers in my field have had to switch from artificial neural networks, which they were keen to describe as a panacea in the 80's, to approaches that were exclusively fueled by mathematical statistics and convex optimization over the last 2 decades. Some of them are even advocating neural networks again! Therefore, academic meetings in machine learning have a very competitive atmosphere, in which researchers enjoy confronting each other and dispute other researchers' findings. These struggles are not what I imagined, as a graduate student, academic life would be. I was expecting a more collaborative and pacified atmosphere. I now recognize that these tensions are, on the contrary, the sign that my field is well alive and growing. Therefore, if you believe your expertise in engineering can help contribute to any of our problems, you are more than welcome to propose new ideas: your approach may well become one of the new hot topics in my field, as long as it works and it is mathematically beautiful!

(准教授 情報学研究科知能情報学専攻)