ICube-MSII seminar: What is the definition of a good machine learning algorithm? After 60 years, is this a closed problem? And if not …

Antoine Cornuéjols (AgroParisTech), will give a talk tuesday 29th may 2018 at 10:00 AM in room A301 of the pole API building in Illkirch.

Title: What is the definition of a good machine learning algorithm? After 60 years, is this a closed problem? And if not …
 
Abstract: Machine Learning has invaded the media sphere thanks to impressive successes such as the victory of a computer Go program against the world champion, or the promise of autonomous vehicles arriving very soon on our roads. Indeed, both the exploitation of massive data and the automatic production of computer code by machines from experience, have put machine learning at the heart of artificial intelligence. Most certainly, then, do we have the answer to the question « what is a good machine learning algorithm? », and what remains is just to adapt the answer to each application field.
Yet, the answer to this question has significantly evolved during the last 60 years, to the extent that papers on machine learning that date back to few decades seem to come from another planet and the concepts therein are indeed no longer taught to our students. And this is not that they appear out of date, but more disturbingly because they seem irrelevant. Have we finally reached the right perspective? Were our precursors wrong? Will our successors quote us in their manuals?
In this presentation, we look at some key moments in the history of machine learning that correspond to turning points in our view of how to evaluate what is a good learning system. And we wonder if we are experiencing another pivotal moment in which are changing again our perspective, the questions we seek to solve in our research, the concepts handled and the way we write our papers.