ScholarWorks@성균관대학교

초록

Virtual drift refers to changes in the input distribution over time while the labels remain unchanged, which can lead to degradation of model performance in long-term deployment. To maintain performance under such input distribution shifts, a model needs to be continuously adapted as new data arrives. However, treating every incoming sample as training data is computationally expensive and often unnecessary, making it essential to identify a small number of samples that are particularly beneficial for adaptation. In this work, we extend the Absolute Learning Progress Gaussian Mixture Model (ALP-GMM) framework, originally developed for task selection in reinforcement learning, to the data sampling setting. Based on this extension, we propose ALP-GMM-based data sampling for drift adaptation. The method estimates per-sample Absolute Learning Progress, which measures how much a sample is expected to improve the model, and then models its distribution using a Gaussian Mixture Model to prioritize samples that are most beneficial for updating the model. Experiments on virtual drift simulated by varying brightness levels in the AI-Hub CCTV dataset show that ALP-GMM-based data sampling improves both adaptation to new data and retention of prior knowledge compared to random sampling baselines. These results demonstrate that the proposed sampling strategy provides an effective and practical approach for maintaining robust performance under virtual drift.

키워드