- 學位論文
基於互補學習之自監督式持續學習框架
Self-Supervised Continual Learning Framework Based on Complementary Learning
摘要
在本論文中，本研究探討了面對持續學習的挑戰，特別是如何在不斷引入新任務和新資料的情況下，使模型能適應新的資料類別，同時保留對先前學到知識的記憶。這在實際應用中尤為困難，因為所面對的資料往往不完整、不精確，且常會出現資料漂移。此外模型必須避免在學習新知識時忘記已習得的知識。為了解決這些問題，本研究提出了一個自監督式的持續學習框架，引入了替代品機制來輔助學習，並加入了一種平衡機制使模型可以公平的學習新舊知識。本研究的系統包括兩個主要模型：快速學習器和慢速學習系統。快速學習器專門處理新的任務，透過自監督學習來快速掌握當前任務的特徵表示。學習完畢後，會將快速學習的結果整合到慢速學習系統中，是透過將這些特徵知識投影到主網路的嵌入空間來完成的。這不僅有助於保留當前任務的關鍵知識，也透過混合使用任務特定資料和替代資料，使模型在無法訪問過去任務資料的情況下，仍能有效回憶過往知識，在最後利用訓練好的主網路利用相似性計算從圖片串流中搜尋與訓練資料有著類似特徵的替代品以解決訓練資料在未來無法使用的問題，這種方法不僅提升了模型對新任務的表現，還增強了其在多變環境下的適應性和泛化能力。
並列摘要
In this thesis, we explore the challenges of continual learning, particularly how to adapt models to new data categories while continually introducing new tasks and data, and retaining memory of previously acquired knowledge. This is particularly difficult in practical applications, as the data encountered is often incomplete, imprecise, and subject to drift. Moreover, the model must avoid forgetting previously learned knowledge while acquiring new information. To address these issues, we propose a self-supervised continual learning framework that introduces a substitution mechanism to aid learning and incorporates a balancing mechanism that allows the model to fairly learn both new and old knowledge. Our system includes two main models: a fast learner and a slow learning system. The fast learner specializes in handling new tasks, rapidly mastering the feature representations of the current task through self-supervised learning. After learning, the results of fast learning are integrated into the slow learning system, which is achieved by projecting the feature knowledge into the embedding space of the main model. This not only helps to retain critical knowledge of the current task but also allows the model to effectively recall past knowledge in the absence of access to past task data, by mixing task-specific data with substitute data. Finally, the trained main model uses similarity calculations to search for substitutes in the image stream that share similar features with the training data, addressing the problem of training data becoming unavailable in the future. This method not only improves the model's performance on new tasks but also enhances its adaptability and generalization capabilities in a changing environment.