Shallow and deep learning : which one should we use? by Michel Verleysen

Machine learning is a branch or artificial intelligence which aims at building models from known data, with few or no hypothesis on the underlying physical phenomenon. Machine learning is nowadays used in many disciplines such as medicine, industrial processes, e-commerce, transport, to cite only a few. Compared to more traditional statistical methods, machine learning is specifically used when the number of features (inputs to the model) is high, i.e. when a high-dimensional regression or classification model of the data has to be designed, and when it is anticipated that the model has to be nonlinear.

The cornerstone of machine learning is the availability of data: it is not difficult to build extremely sophisticated nonlinear models of the relation between a high number of features and output(s), but the lack of hypotheses on the underlying phenomenon has to be compensated by a lot of data (and sometimes really a lot of…). In some domains it is easy to collect huge amounts of data, in other domains it is far more difficult (patient records for example). The difficulty is not to build a model that fits well data, but to build a model that does not overfit the data especially when data are scarse, which is unavoidable in high-dimensional spaces.

This talk will present the main lines of the state-of-the-art in machine learning, including but not limited to deep learning, the main challenges, and opportunities for research purposes and applications.