“Logical, epistemological and ethical limitations of machine learning”
The growing echo of Big Data and Deep Learning in the last years, poorly mask their intrinsic limitations. They are adorned with all the virtues to the point that it is assumed that they will transform humanity to make it more just and above all more perfect, and perhaps help it to reach if not immortality, at least a greater longevity. Yet, while these techniques are and will be important in many sectors, such as in the medical field, to help diagnose diseases, or in the social field, to predict and rationalize certain choices, they are based on induction, that is to say on the reasoning that goes from the particular to the general. Consequently, they are subject to the logical limitations of any induction that we will try to recall here. Moreover, as we will show, these techniques make it possible to detect correlations that do not all correspond to causal relationships, and that sometimes prove to be misleading. Finally, we will see that the misuse of decision-making based on machine learning can have predatory effects on society because, far from being objective, the choices include biases.
