La vraie utilité de l’adaptation des paramètres des couches de normalisation d’un CNN en Fully Test Time Adaptation

Abstract

With the promising results of Fully Test Time Adaptation methods, the computer vision community has increased its research interest in thoroughly investigating this approach. The fundamental feature of a Fully Test Time Adaptation method is that it foregoes updating a neural network to a data shift during the learning phase, limiting the adaptation to the inference phase on each new test batch. Examining the literature, we found that most works studying this approach focus on updating batch normalization parameters. While the obtained results are encouraging, one interesting question arises : do we really need to update all the batch normalization parameters to increase the model performance in inference ? To answer this question, we propose to study the real usefulness of adapting batch normalization parameters in Fully Test Time Adaptation. In this context, we have carried out a comparative study to illustrate how FTTA performs on models built for head and neck tumor segmentation from CT-Scan and PET-Scan under various settings. Our findings show that the parameters to be adapted in FTTA represent the most critical factor to be considered. It also shows that the gain obtained by updating the batch norm parameters (alpha and beta) is negligible compared to the improvement obtained by setting the right batch norm statistical parameters that correspond to the actual batch test. This casts doubt on the usefulness of updating the batch normalization’s scale and bias parameters often used in FTTA in the literature. For more generalizability, we extended our study to a second scenario related to natural image classification using the CIFAR dataset. The new results confirm our first findings, strengthening our doubt about the utility of adapting the batch normalization’s parameters in FTTA. The code of this project can be found in the following GitHub repository : https://github.com/ghassenbaklouti/FTTA-For-HNTS.

Date	19 Sept 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Ismail Ben Ayed (Supervisor) & Houda Bahig (Co-supervisor)