Rapport de synthèse : Un cadre simple pour l'apprentissage contrastif des représentations visuelles
Michel, Théo, Widad Zizouan · HAL (Le Centre pour la Communication Scientifique Directe) · 2022
This is a short article trying to reproduce the some of the findings of the SimCLR paper [1]. The SimCLR paper was published at the ICML conference in 2020 by the Google AI team. The following experiment has been done with different hyper-parameters and with sometimes different data sets than the original paper, but still managed to observe around the same relative effectiveness to other supervised and self-supervised methods trained in the same conditions. The results found here are in no way an affirmation of the original paper, just a hint at the possibility of using those same methods with smaller training requirements. This was written in the context of a second year introduction to doctoral research at Enseirb-Matmeca by computer science students.