基於難易度感知之混合數據增強對於經驗回放持續學習之方法

Abstract

深度神經網路在經過持續學習會遭遇到災難性遺忘的問題，學習新知會影響舊知的保存，使得模型在舊任務的表現嚴重下滑，而經驗回放是透過儲存樣本，在接收到新任務的資料時同時進行回放訓練，為解決持續學習災難性遺忘的最有效的方法之一，但由於儲存空間的有限，每一個過去的任務只會儲存少量的資料，會遇到新舊資料不平衡的問題，因此本文提出透過混和數據增強的方式增加過去回放樣本的多樣性，另外基於困難度決定混和參數，消除由於隨機混和所造成的困難樣本模糊化的現象，所提出的方法可以直接應用在經驗回放之方法上，我們將方法實作在ER、DER、以及DER++上，並在split cifar-10、split cifar-100、split mini-ImageNet數據集上檢測我們的方法，實驗結果顯示，所提出的方法在沒有增加過多的計算資源下可以有效的提高平均正確率，並減輕了遺忘程度。

Deep neural networks suffer from the issue of catastrophic forgetting after continual learning, causing a sudden deterioration in performance when training on new tasks. Replay-based methods, which is one of the most effective solutions, alleviate catastrophic forgetting by replaying the subset of past data stored in memory buffer. However, due to the limited storage space, a small amount of past data can be stored, and there will form a data imbalance situation between old and new tasks. Hence, in this work, we tried to increase the diversity of past samples by mixup. In addition, we propose difficulty-aware mixup approach that modifies the mixing coefficient according to the distance between output logits and ground truth labels to reduce the ambiguity of hard examples. We implement our method on ER, DER, and DER++, and test it on split-CIFAR10, split-CIFAR100, and split-miniImagenet. The experimental result shows that the proposed method can effectively improve the average accuracy and reduce the forgetting without adding too many computing resources.