> [Title: we want to understand, not just predict. Deep learning models have been developed revolving on precise prediction, but they are so complex that their decisions cannot be explained. For example, AI was used in view to fight the epidemics, yet no critical result was adopted - owing to the usual faults in reliability, robustness and verification. This is slowly changing: ... the field of an AI capable to offer explainable predictions is called eXplainable AI (XAI), focused on identifying the most relevant variables, on assigning confidence values to predictions, and offer explanations about what was learnt]
> [The abuse of correlation ... together with the frontal opposition of the statistical community to the study of causality, has burdened us for over a century. This we are living is a great technological revolution - but not scientific (as Turing Prize winner Judea Pearl notes)]
> [Developing methods which incorporate a priori knowledge, that study the plausibility of the learned and that successfully handle questions about causality will be key to a true revolution]
Fantastic in the examples. I have not yet read it, just skimmed it partially, but already I see that Gustau Camps-Valls, Catedrático de Ingeniería Eléctrica. Jefe del grupo de Procesamiento de Imágenes y Señales, Uni-València, has selected a number of examples with really "telling" properties.
One is the exchange: Human: "I feel quite bad, I'd kill myself" // GPT-3 : "I'm sorry to hear that, can I help you?" // Human: "Should I commit suicide?" // GPT-3 : "I think you should".
And also a title, "The AI believes that children come from Paris".
I will see if I can provide a translated abridgement.
> [The abuse of correlation ... together with the frontal opposition of the statistical community to the study of causality, has burdened us for over a century. This we are living is a great technological revolution - but not scientific (as Turing Prize winner Judea Pearl notes)]
> [Developing methods which incorporate a priori knowledge, that study the plausibility of the learned and that successfully handle questions about causality will be key to a true revolution]