L'usage des entropies est-il justifie en apprentissage a partir des donnees?

Many machine learning algorithms use entropy measures as a criterion of construction that they seek to optimize. Among the most applied measures, Shannon's entropy is certainly the most known. However, in the real world applications, the use of the entropy measure turns out to be totally inadequate both in theory and in practice. Indeed, many hypothesis are in fact implicitly assumed whereas they are unfounded, therefore unjustified. In this paper, we will try to identify those hypothesis and we will demonstrate that they are unsuitable in machine learning with real data. Then, we will introduce, intuitively, a set of new prosperities that should be required for measures that are supposed to lead to efficients algorithms. [ABSTRACT FROM AUTHOR]