Leveraging Proxy of Training Data for Test-Time Adaptation

Authors: Juwon Kang, Nayeong Kim, Donghyeon Kwon, Jungseul Ok, Suha Kwak

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On four public benchmarks, our method outperforms the state-of-the-art ones at remarkably less computation and memory. [...] 4. Experiments
Researcher Affiliation Academia 1Department of Computer Science and Engineering, POSTECH, Pohang, Korea 2Graduate school of Artificial Intelligence, POSTECH, Pohang, Korea. Correspondence to: Suha Kwak <suha.kwak@postech.ac.kr>.
Pseudocode No The paper describes its methods in prose but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement or link for open-source code for the methodology.
Open Datasets Yes The proposed method is first evaluated on three benchmarks for common image corruptions (CIFAR10-C, CIFAR100-C, and Tiny Image Net-C (Hendrycks & Dietterich, 2019)) and one benchmark for synthetic-to-real adaptation (Vis DA-C (Peng et al., 2017)).
Dataset Splits Yes Table 3. Classification accuracy (%) on Vis DA-C train val. All methods use Res Net-101 backbone.
Hardware Specification No No specific hardware details (e.g., GPU models, CPU types, or cloud instance specifications) used for running experiments are explicitly mentioned in the paper.
Software Dependencies No The paper mentions optimizers (Adam, SGD) but does not provide specific version numbers for any software libraries or dependencies used for the experiments.
Experiment Setup Yes For style-normalized dataset condensation, we set the number of synthetic images per class to 10 for CIFAR and Tiny Image Net and to 50 for Vis DA-C. For optimization in condensation, we initialize the synthetic images by random noise and optimize them with SGD. [...] For optimization in test time, we adopt Adam (Kingma & Ba, 2015) for the common image corruption benchmark and SGD for Vis DA-C. The balancing parameter λ is set to 0.1 for the corruption benchmark and 1.0 for the other. For temperature control, we set τ1 to 0.1 and τ2 to the inverse of the square root of the number of classes for each dataset.