Bremen 2017 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

AGPhil: Arbeitsgruppe Philosophie der Physik

AGPhil 3: Symposium: Epistemology of Big Data in Physics I

AGPhil 3.1: Hauptvortrag

Donnerstag, 16. März 2017, 13:30–14:15, GW2 B2900

Data-driven hypothesis generation using deep neural nets — •Balázs Kegl — CNRS / Université Paris-Saclay

Generating and testing a large number of low-probability hypotheses in certain scientific fields lead to the so called p value controversy. From the point of view of hard sciences this seems as an abnormal misuse of the scientific method. In the first part of the talk I will argue that the scientific method, as it is understood today, does not prevent these aberrations. In tomorrow's world where computational tools can generate scientific hypotheses automatically, fixing this issue is of uttermost importance. Solving this problem will require putting hypothesis generation back into the center of the scientific method.

The goal of computational creativity is to design methods that can generate valuable novelty. One major debate within this community is whether generation is mostly random (only the evaluation process has a strong notion of value of novelty), or we should include knowledge already in the generative process. I will show how this issue is related to the p value controversy and automatic hypothesis generation. I will present a constructive framework in which data- and knowledge-driven novelty generation can be studied and evaluated. I will finish the talk by showing some of our latest results using deep neural nets as the knowledge representation and novelty generation engine.