PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptation

Authors: Sarthak Kumar Maharana, Baoming Zhang, Yunhui Guo

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive image classification experiments on CIFAR-10C, CIFAR100C, and Image Net-C, demonstrating the superior efficacy of our method compared to prior approaches.
Researcher Affiliation Academia Sarthak Kumar Maharana, Baoming Zhang, Yunhui Guo The University of Texas at Dallas, Richardson, USA EMAIL
Pseudocode No The paper describes the methodology using equations and prose, but does not contain any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code https://github.com/sarthaxxxxx/PALM
Open Datasets Yes Following the standard benchmarks set by (Wang et al. 2022), we evaluate our proposed method on CIFAR-10C, CIFAR-100C, and Image Net-C, based on image corruption schemes as set in (Hendrycks and Dietterich 2019).
Dataset Splits Yes Following the standard benchmarks set by (Wang et al. 2022), we evaluate our proposed method on CIFAR-10C, CIFAR-100C, and Image Net-C, based on image corruption schemes as set in (Hendrycks and Dietterich 2019). ... To maintain fairness regarding the source models, we employ Wide Res Net-28 ... Res Ne Xt-29 ... and Res Net-50 ..., all available on Robust Bench (Croce et al. 2020).
Hardware Specification Yes With a single NVIDIA A5000 GPU, PALM incurs a slightly higher adaptation time/batch, due to the two-stage proposed method.
Software Dependencies No The paper mentions using an Adam optimizer and specific model architectures like Wide Res Net-28, Res Ne Xt-29, and Res Net-50, but does not provide specific version numbers for software libraries or dependencies.
Experiment Setup Yes We optimize using an Adam optimizer, setting base learning rates (κ) to 5e-4 for CIFAR-10C and CIFAR100C, and 5e-5 for Image Net-C. For balancing parameter sensitivities, we set α to 0.5, 0.9, and 0.5 respectively, with temperature coefficients T set to 50, 100, and 1000 respectively. We set η to 1, 0.5, and 0.3, and λ to 0.01 throughout. Batch sizes are set to 200, 200, and 64 for each dataset, following Co TTA, and results are only reported for the highest severity level of 5 for each task.