Stationary Latent Weight Inference for Unreliable Observations from Online Test-Time Adaptation

Authors: Jae-Hong Lee, Joon-Hyuk Chang

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

Reproducibility Variable Result LLM Response
Research Type Experimental The efficacy of SLWI is demonstrated through various experimental setups, showcasing its superior performance in diverse distribution shift scenarios.We evaluate the proposed framework across a multitude of distribution-shift scenarios and datasets, which show significant performance improvements compared with current state-of-the-art methods.
Researcher Affiliation Academia Jae-Hong Lee 1 Joon-Hyuk Chang 1 1Department of Electronic Engineering, Hanyang University, Seoul, Republic of Korea.
Pseudocode Yes Algorithm 1 General OTTA Framework Algorithm 2 Stationary Latent Weight Inference Framework Algorithm 3 Stationary Latent Weight Inference
Open Source Code No The paper does not provide explicit open-source code for the proposed SLWI framework.
Open Datasets Yes We conducted experiments on two standard datasets, Image Net-C (Hendrycks & Dietterich, 2019a) and D109 (Marsden et al., 2023), which represent corruption and natural distribution shifts that occur in the wild-world (Niu et al., 2023). Image Net (Deng et al., 2009) contains 1,281,167 training images and 50,000 test images.
Dataset Splits Yes Image Net (Deng et al., 2009) contains 1,281,167 training images and 50,000 test images. The average and standard deviation of error rates for random seeds 0-4 were used as the evaluation metrics.
Hardware Specification Yes Our experiments were conducted using a single NVIDIA Ge Force RTX 3090 GPU
Software Dependencies No The paper mentions the use of data2vec, Vision Transformer, Swin Transformer, and SGD optimizer, but does not specify version numbers for any software libraries or dependencies.
Experiment Setup Yes We set the batch size to 64, the learning rate to 0.000014, and trained the models using an SGD optimizer. Unless specifically mentioned, the ROID objective function was primarily adopted for our framework. The SLWI parameters, (a, q), were set to (0.99, 0.001), and the strict Kalman gain b for the transfer model was set to the same value as a, that is 0.99. The degree of backtracking α was set to 1.4.