任務順序於持續學習之影響

Task Order Effect in Continual Learning

屈佳宏(Jia-Hong Qu)

指導教授 ： 陳銘憲

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202201598

摘要

在持續學習的文獻中，由於非平穩的數據分佈變化，大多數研究致力於克服災難性遺忘的問題，這對於人工神經網絡而言是個嚴重問題。然而，大多數的研究主要致力在持續學習「演算法」的部分，而任務順序之於機器學習模型的影響卻鮮少受到關注。在這篇論文中，我們研究了持續學習中的任務順序效應，並發現任務順序的選擇可能會影響最終的模型性能。因此，我們設計了一個名為啟發式任務選擇 的框架，它可以加入現有算法中，並在漸近多項式的時間內有效地選擇下一個任務。此外，我們利用幾種啟發式策略作為我們的選擇標準，並顯示這種基於特徵空間中的估計方法能作為有效性的指標，藉此可以很好地組織分類任務的任務順序。而為了評估具有大量任務的基準，我們利用符號檢定，一種可用於測量統計量的無母數或無分佈檢定，來評估我們所選擇的任務順序。實驗結果揭示了我們的方法在幾個基準上的有效性。

關鍵字

持續學習 ； 終身學習 ； 任務順序效應 ； 領域增量式學習 ； 數據分佈變化

並列摘要

In continual learning literature, due to the non-stationary data distribution shifts, most research is dedicated to overcoming the catastrophic forgetting problem, which is a severe issue for artificial neural networks. Nevertheless, most works mainly focused on the algorithm part of continual learning, while the task order presented to machine learning models received little attention. In this work, we investigate the task order effect in continual learning and show that the choice of task order can impact the final performance. Therefore, we design the framework named Heuristic Task Selection, which can be plugged into existing algorithms and efficiently select the next task. Furthermore, we utilize several heuristics as our selection criterion and show that such feature-based metrics can be good indicators to well organize the task order for classification tasks. For evaluation, we exploit the sign test, a non-parametric or distribution-free test, which can be applied to measure the location of a statistic and help assess how performant the task order we select for the benchmarks with enormous tasks. The results reveal the efficacy of our method on several commonly used benchmarks.

並列關鍵字

Continual Learning ； Lifelong Learning ； Task Order Effect ； Domain Incremental Learning ； Data Distribution Shift

參考文獻

[1] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6430–6439, 2019.

[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366–3375, 2017.

[3] Yoshua Bengio,Jérôme Louradour,Ronan Collobert,and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48, 2009.

[4] Rich Caruana. Multitask learning. Machine learning, 28(1):41–75, 1997.

[5] Arslan Chaudhry, Marc’Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420, 2018.