2025-05-13
【學術亮點-頂級會議論文】在資源有限環境中逐步適應的持續學習方法
字體大小
小
中
大
【學術亮點-頂級會議論文】在資源有限環境中逐步適應的持續學習方法
AI core Technology: Advanced Research and Resource Integration Platform or AI Technology【Department of Computer Science and Engineering / Yao-Chung Fan / Professor】
核心技術:AI核心技術之進階研究與資源整合平台【資訊工程學系范耀中教授】
上架日期:2025
AI core Technology: Advanced Research and Resource Integration Platform or AI Technology【Department of Computer Science and Engineering / Yao-Chung Fan / Professor】
核心技術:AI核心技術之進階研究與資源整合平台【資訊工程學系范耀中教授】
論文篇名 | 英文:GAIN: Gradual Adaptation for Continual Learning in Low-resource Environments 中文:在資源有限環境中逐步適應的持續學習方法 |
期刊名稱 | 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing (WWW 2025) (指標清單會議) |
發表年份, 卷數, 起迄頁數 | In Proceedings of the ACM Web Conference 2025 (WWW Companion ’25), April 28-May 2, 2025, Sydney, NSW, Australia. ACM, New York, NY, USA, 5 pages. |
作者 | Shih-Wei Guo; Yao-Chung Fan(范耀中)∗ |
DOI | 10.1145/3701716.3715499 |
中文摘要 | 持續學習(Continual Learning, CL)在資源有限的任務中面臨災難性遺忘與資料稀缺等挑戰,這限制了模型對逐漸變化任務的適應能力。我們提出了一種針對低資源任務的通用持續學習方法,名為 GAIN,以應對這些挑戰。該方法透過兩個策略來處理跨領域的持續學習問題:(1)「逐步適應模組」(Gradual Adaptation Module),此模組逐步堆疊輕量化的適配器,以有效保留先前任務的知識並適應新任務,從而減輕災難性遺忘;(2) 瓶頸層大小調整策略,可提升低資源場景下的學習效率。最後,我們在四個資料集上進行了大量實驗,以驗證 GAIN 的有效性與穩健性,結果顯示其能減輕災難性遺忘,並在各類任務中提升學習效率。 |
英文摘要 | Continual Learning (CL) faces challenges such as catastrophic forgetting and data scarcity in low-resource tasks, which limit the model's ability to adapt to gradually changing tasks. We propose a general CL method for low-resource tasks to address these challenges called GAIN. This method addresses cross-domain continual learning through two approaches: (i) the Gradual Adaptation Module, which incrementally stacks lightweight adapters to effectively retain knowledge from previous tasks and adapt to new ones, thereby mitigating catastrophic forgetting, and (ii) a bottleneck layer size tuning strategy to improve learning efficiency in low-resource scenarios. Finally, we have extensive experiments on four datasets to validate the effectiveness and robustness of GAIN, showing that it alleviates catastrophic forgetting and enhances learning efficiency across various task types. |
發表成果與AI計畫研究主題相關性 | 農業現場資料昂貴、收集難,尤其是精細標註如作物疾病類別。 GAIN 提出瓶頸層大小調整策略,可提升在資料稀缺情況下的學習效率,讓農業AI在低算力裝置或遠端場域仍能有效運行。 |