Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment

Authors: Zhen-Yu Zhang, Zhiyu Xie, Huaxiu Yao, Masashi Sugiyama

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on both benchmark datasets and a real-world application validate the effectiveness and adaptability of our proposed algorithm. Through benchmark experiments and a real-world application, we demonstrate the effectiveness and efficiency of our method in improving performance under non-stationary environments.
Researcher Affiliation Academia Zhen-Yu Zhang Center for Advanced Intelligence Project, RIKEN zhen-yu.zhang@riken.jp Zhiyu Xie Stanford University zhiyuxie@stanford.edu Huaxiu Yao UNC-Chapel Hill huaxiu@cs.unc.edu Masashi Sugiyama Center for Advanced Intelligence Project, RIKEN Graduate School of Frontier Sciences, The University of Tokyo sugi@k.u-tokyo.ac.jp
Pseudocode Yes Algorithm 1 Adaptive Representation Alignment 1: Initialization: i [K], ϕi t = ϕ0 2: for t = 1 to T do 3: for i = 0 to K do 4: if 2i mod t == 0 then 5: set ϕi t = ϕ0, Ri t = 0, Ci t = 0 6: end if 7: end for 8: update base learners by Eqn. (4) and update weight pt K according to Eqn. (6) 9: combine base learners according to Eqn. (5) 10: update classifier module according to Eqn. (7) 11: end for
Open Source Code Yes Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: All the details of the data can be found in A.1. Codes can be found in Supplementary Material.
Open Datasets Yes We evaluate the proposed Ada-Re Align algorithm on two large-scale benchmark datasets: CIFAR10C and Image Net C. We further evaluate the proposed algorithm on a real-world wildlife species classification task using the i Wild Cam dataset [3]
Dataset Splits No The paper mentions training on 'clean data' and testing on 'corrupted data streams' but does not specify details for a validation split (percentages, counts, or how it was used) nor does it cite a standard split that includes validation.
Hardware Specification Yes We use the following compute resource configuration: 2 Xeon Gold 6242R with a base frequency of 3.1 GHz, 8 Ge Force 3090 with 24GB VRAM, and a total of 768GB RAM.
Software Dependencies No The operating system employed is Ubuntu 20.04. The paper does not provide specific version numbers for programming languages (e.g., Python) or libraries (e.g., PyTorch, scikit-learn) that would be needed for replication.
Experiment Setup Yes For our proposed algorithm, we use SGD as the update rule, with a momentum of 0.9 and a learning rate of 0.0005. The optimizer employed is SGD, with a momentum of 0.9 and a batch size of 32 for rounds where the number of data data exceeds 32. ... The learning rate is set to 0.0005.