TEA: Test-Time Energy Adaptation

The appendix contains the following sections: (1) Additional Experiments and Analyses (Sec. 7): -Detailed Results for Energy Reduction (Sec. 7.1) -Detailed Results for Image Corruption (Sec. 7.2) -Hyper-parameters Sensitivity (Sec. 7.3) (2) Detailed Settings (Sec. 8): -Datasets(Sec. 8.1) -Evaluation Metrics(Sec. 8.2) -Hyper-parameters(Sec. 8.3) -Computing Resources (Sec. 8.4) (3) Limitations and Future Explorations (Sec. 9). Additional Experiments Detailed Results for Energy Reduction This section serves as an extension of the energy analysis (Sec. 4.3.1) in the main text, presenting the relationship between TEA's energy reduction and the enhancement of generalizability across all types of corruption. The detailed results are shown in Figs. 7 and 8 , where each corruption type is analyzed at five levels of severity, with the analysis examining the correlation between the extent of energy reduction and performance improvements, both before and after adaptation, as severity levels increase. In our experiments, TEA generally reduced energy and enhanced generalization across various corruptions. Yet, for mild corruptions like "Brightness" at level one, i.e., the mildest in CIFAR-10-C, generalization did not improve and occasionally deteriorated slightly. Correspondingly, energy did not decrease and even increased marginally. These outcomes indicate a strong correlation between generalizability enhancement and energy reduction. However, it is possible that our method may not reduce energy as anticipated for distributions with some less severe corruption types. This may be attributed to these distributions being closely aligned with the original, already at a low energy state. The uniform hyperparameters used in our adaptation may not be optimal for such cases. Addressing this discrepancy will be a priority in future research.