A Layer Selection Approach to Test Time Adaptation

Authors: Sabyasachi Sahoo, Mostafa ElAraby, Jonas Ngnawe, Yann Batiste Pequignot, Frédéric Precioso, Christian Gagné

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments, we demonstrate that the proposed layer selection framework improves the performance of existing TTA approaches across multiple datasets, domain shifts, model architectures, and TTA losses.
Researcher Affiliation Academia 1IID, Universit e Laval 2Mila 3Universit e de Montr eal 4Universit e Cote d Azur, CNRS, INRIA, I3S, Maasai 5Canada CIFAR AI Chair EMAIL
Pseudocode No The paper describes the proposed approach using mathematical equations and descriptive text, but it does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block with structured steps.
Open Source Code No The paper does not contain any explicit statements about releasing code, nor does it provide links to any code repositories or mention code in supplementary materials.
Open Datasets Yes This section compares our proposed approaches with existing baselines on Domainbed (Gulrajani and Lopez-Paz 2021), a popular benchmark with large single distribution shifts, and Continual TTA, a popular benchmark with multiple distribution shifts. TTA losses Two popular TTA losses are considered: Pseudo-Labeling (PL) (Lee et al. 2013) and SHOT (Liang, Hu, and Feng 2020).
Dataset Splits Yes For the experiments on Domainbed, we follow the evaluation protocol as described in Iwasawa and Matsuo (2021), including dataset splits for the following four datasets: PACS (Li et al. 2017), VLCS (Fang, Xu, and Rockmore 2013), Terra Incognita (Beery, Van Horn, and Perona 2018), and Office-Home (Venkateswara et al. 2017).
Hardware Specification No Computations were made on the cedar, and beluga supercomputers, managed by Calcul Qu ebec and the Digital Research Alliance of Canada (Alliance).
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup Yes Implementational details We report results for GALA with window size of 20 and selection threshold of 0.75 with single-layer granularity. It appears that GALA is not overly sensitive to hyperparameters, and those values work well overall see Sec. 5 for more discussion on hyperparameter values and the design choices.