一种基于不确定性校正的可信连续学习方法

doi:10.19678/j.issn.1000-3428.0070658

摘要/Abstract

摘要：

连续学习是一种能在连续的数据流中训练模型的深度学习范式, 适合日益开放复杂的智能应用场景。连续学习最大的挑战是"灾难性遗忘", 即模型在学习新知识后会遗忘过去学习的知识。现有的连续学习忽视了不确定性对模型训练的影响, 且大多数方法都聚焦于连续学习后续阶段的改进, 对模型初始阶段研究较少。提出一种基于不确定性校正的可信连续学习方法, 通过在初始阶段约束模型输出的不确定性来弥补模型参数漂移带来的分类误差, 从而缓解"灾难性遗忘"。该方法可以与其他连续学习方法相结合并形成改进模型, 具有较强的通用性。对3种经典的连续学习方法进行改进, 实验结果表明所提方法均能有效提高原始模型的性能, 在2个数据集上平均分类准确率(ACA)提升1.2~19.1百分点。除此之外, 引入期望校准误差(ECE)评判连续学习模型的可靠性, 对比原始模型, 基于所提方法的改进模型具有更低的ECE, 这证明基于该方法的改进模型更加可信。

关键词: 连续学习, 不确定性, 校正, 信息熵, 知识蒸馏

Abstract:

As a deep learning paradigm that can be trained in continuous data streams, continuous learning is suitable for increasingly open and complex intelligent application scenarios. The main challenge in incremental learning is catastrophic forgetting, which refers to a precipitous drop in performance on previously learned tasks after learning a new task. State-of-the-art studies on continuous learning ignore the impact of uncertainty on model training. In addition, existing studies mainly focus on mitigating forgetting in the phases after the initial phase, whereas the role of the initial phase is largely neglected. Motivated by this, this study proposes a trusted continuous learning method based on uncertainty correction that constrains the uncertainty of a model at the initial stage. Thus, this constraint can alleviate errors caused by model parameter drift, and catastrophic forgetting can be relieved. The proposed method can be combined with other continuous learning methods; therefore, it is fairly universal; for example, this study improves three traditional continuous learning models using the proposed method, and the experimental results show that the improved models outperform the original ones, with the Average Classification Accuracy (ACA) improving by 1.2 to 19.1 percentage points on two datasets. The Expected Calibration Error (ECE) is used to evaluate the reliability of the models. The experimental results show that the improved models have a lower ECE, which proves that the proposed method improves the reliability of the original models.

Key words: continuous learning, uncertainty, correction, information entropy, knowledge distillation

沈勤丰, 黄璐瑶. 一种基于不确定性校正的可信连续学习方法[J]. 计算机工程, 2025, 51(12): 56-67.

SHEN Qinfeng, HUANG Luyao. A Trusted Continuous Learning Method Based on Uncertainty Correction[J]. Computer Engineering, 2025, 51(12): 56-67.

https://www.ecice06.com/CN/Y2025/V51/I12/56

图/表 17

图1 UC-RCL结构

Fig.1 UC-RCL structure

图2 连续学习基本框架

Fig.2 The basic framework for continuous learning

图3 CIFAR100数据集部分图像

Fig.3 Partial images of CIFAR100 dataset

图4 SVHN数据集部分图像

Fig.4 Partial images of SVHN dataset

图5 CIFAR100_5/20各阶段ACA折线图

Fig.5 The ACA line chart of CIFAR100_ 5/20 in each stage

图6 CIFAR100_10/10各阶段ACA折线图

Fig.6 The ACA line chart of CIFAR100_ 10/10 in each stage

图7 SVHN_5/2各阶段ACA折线图

Fig.7 The ACA line chart of SVHN_5/2 in each stage

图8 LwF与LwF-UC的ECE对比图

Fig.8 Comparison chart of ECE between LwF and LwF-UC

图9 模型输出的熵值对比(LwF和LwF-UC)

Fig.9 Comparison of entropy values of model outputs (LwF and LwF-UC)

图10 阈值u_th对LwF-UC性能的影响

Fig.10 Effect of u_th on the performance of LwF-UC

参考文献 41

1	张东阳, 陆子轩, 刘军民, 等. 深度模型的持续学习综述: 理论、方法和应用. 电子与信息学报, 2024, 46 (10): 3849- 3878.
	ZHANG D Y , LU Z X , LIU J M , et al. A survey on continual learning of deep models: theories, methods, and applications. Journal of Electronics & Information Technology, 2024, 46 (10): 3849- 3878.
2	CHEN Z Y , LIU B . Lifelong machine learning. Synthesis Lectures on Artificial Intelligence and Machine Learning, 2016, 10 (3): 1- 145. doi: 10.1007/978-3-031-01581-6
3	KUDITHIPUDI D , AGUILAR-SIMON M , BABB J , et al. Biological underpinnings for lifelong learning machines. Nature Machine Intelligence, 2022, 4 (3): 196- 210. doi: 10.1038/s42256-022-00452-0
4	韩亚楠, 刘建伟, 罗雄麟. 连续学习研究进展. 计算机研究与发展, 2022, 59 (6): 1213- 1239.
	HAN Y N , LIU J W , LUO X L . Research progress of continual learning. Journal of Computer Research and Development, 2022, 59 (6): 1213- 1239.
5	HINTON G , VINYALS O , DEAN J . Distilling the knowledge in a neural network. Computer Science, 2015, 14 (7): 38- 39.
6	LI Z Z , HOIEM D . Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40 (12): 2935- 2947. doi: 10.1109/TPAMI.2017.2773081
7	CASTRO F M, MARÍN-JIMÉNEZ M J, GUIL N, et al. End-to-end incremental learning[C]//Proceedings of the European Conference on Computer Vision (ECCV). Berlin, Germany: Springer, 2018: 233-248.
8	REBUFFI S A, KOLESNIKOV A, SPERL G, et al. iCaRL: incremental classifier and representation learning[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, USA: IEEE Press, 2017: 5533-5542.
9	MASANA M , LIU X L , TWARDOWSKI B , et al. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45 (5): 5513- 5533. doi: 10.1109/TPAMI.2022.3213473
10	LEE K, LEE K, SHIN J, et al. Overcoming catastrophic forgetting with unlabeled data in the wild[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Seoul, Korea: IEEE Press, 2019: 312-321.
11	ZHANG J T, ZHANG J, GHOSH S, et al. Class-incremental learning via deep model consolidation[C]//Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV). Snowmass Village, USA: IEEE Press, 2020: 1120-1129.
12	HOU S H, PAN X Y, LOY C C, et al. Learning a unified classifier incrementally via rebalancing[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE Press, 2019: 831-839.
13	KANG M, PARK J, HAN B. Class-incremental learning by knowledge distillation with adaptive feature consolidation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, USA: IEEE Press, 2022: 16050-16059.
14	LIU Y , HONG X P , TAO X Y , et al. Model behavior preserving for class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems, 2023, 34 (10): 7529- 7540. doi: 10.1109/TNNLS.2022.3144183
15	PENG C, ZHAO K, WANG T R, et al. Few-shot class-incremental learning from an open-set perspective[C]//Proceedings of the 17th European Conference on Computer Vision (ECCV). Berlin, Germany: Springer, 2022: 382-397.
16	XUE H, AN Y X, QIN Y C, et al. Towards few-shot learning in the open world: a review and beyond[EB/OL]. [2023-12-19]. https://arxiv.org/pdf/2408.09722.
17	ZHAO L L, LU J, XU Y L, et al. Few-shot class-incremental learning via class-aware bilateral distillation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Vancouver, Canada: IEEE Press, 2023: 11838-11847.
18	LIN J H , WU Z H , LIN W F , et al. M2SD: multiple mixing self-distillation for few-shot class-incremental learning. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38 (4): 3422- 3431. doi: 10.1609/aaai.v38i4.28129
19	KIRKPATRICK J , PASCANU R , RABINOWITZ N , et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America, 2017, 114 (13): 3521- 3526.
20	WU Y C, HUANG L K, WANG R Z, et al. Meta continual learning revisited: implicitly enhancing online hessian approximation via variance reduction[C]//Proceedings of International Conference on Learning Representations. Vienna, Austria: [s. n.], 2024: 1-30.
21	LIN H W, ZHANG B Q, FENG S S, et al. PCR: proxy-based contrastive replay for online class-incremental continual learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Vancouver, Canada: IEEE Press, 2023: 24246-24255.
22	OSTAPENKO O, PUSCAS M, KLEIN T, et al. Learning to remember: a synaptic plasticity driven framework for continual learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE Press, 2019: 11313-11321.
23	WANG L Y, YANG K, LI C X, et al. ORDisCo: effective and efficient usage of incremental unlabeled data for semi-supervised continual learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, USA: IEEE Press, 2021: 5379-5388.
24	ZHU F, ZHANG X Y, WANG C, et al. Prototype augmentation and self-supervision for incremental learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, USA: IEEE Press, 2021: 5867-5876.
25	朱飞, 张煦尧, 刘成林. 类别增量学习研究进展和性能评价. 自动化学报, 2023, 49 (3): 1- 26.
	ZHU F , ZHANG X Y , LIU C L . Class incremental learning: a review and performance evaluation. Acta Automatica Sinica, 2023, 49 (3): 1- 26.
26	LI X L, ZHOU Y B, WU T F, et al. Learn to grow: a continual structure learning framework for overcoming catastrophic forgetting[C]//Proceedings of the International Conference on Machine Learning. New York, USA: ACM Press, 2019: 3925-3934.
27	YAN S P, XIE J W, HE X M. DER: dynamically expandable representation for class incremental learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, USA: IEEE Press, 2021: 3013-3022.
28	ZHOU D W, WANG Q W, YE H J, et al. A model or 603 exemplars: towards memory-efficient class-incremental learning[EB/OL]. [2023-12-19]. https://arxiv.org/pdf/2205.13218.pdf.
29	ZHAO H B , FU Y J , KANG M T , et al. MgSvF: multi-grained slow versus fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46 (3): 1576- 1588. doi: 10.1109/TPAMI.2021.3133897
30	DOUILLARD A, RAME A, COUAIRON G, et al. DyTox: transformers for continual learning with DYnamic TOken eXpansion[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, USA: IEEE Press, 2022: 9275-9285.
31	WANG Z F, ZHANG Z Z, LEE C Y, et al. Learning to prompt for continual learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, USA: IEEE Press, 2022: 139-149.
32	MITTAL S, GALESSO S, BROX T. Essentials for class incremental learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Nashville, USA: IEEE Press, 2021: 3508-3517.
33	CUBUK E D, ZOPH B, MANE D, et al. AutoAugment: learning augmentation strategies from data[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE Press, 2019: 113-123.
34	FURLANELLO T, LIPTON Z, TSCHANNEN M, et al. Born again neural networks[C]//Proceedings of the International Conference on Machine Learning. New York, USA: ACM Press, 2018: 1607-1616.
35	SHI Y J, ZHOU K Q, LIANG J, et al. Mimicking the oracle: an initial phase decorrelation approach for class incremental learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, USA: IEEE Press, 2022: 16701-16710.
36	GUO C, PLEISS G, SUN Y, et al. On calibration of modern neural networks[C]//Proceedings of International Conference on Machine Learning. New York, USA: ACM Press, 2017: 1321-1330.
37	KRISHNAN R , TICKOO O . Improving model calibration with accuracy versus uncertainty optimization. Neural Information Processing Systems, 2020, 33, 18237- 18248.
38	KRIZHEVSKY A. Learning multiple layers of features from tiny images[M]//ASHERSON R A. Handbook of systemic autoimmune diseases. [S. l.]: Elsevier, 2009: 1-10.
39	NETZER Y, WANG T, COATES A, et al. Reading digits in natural images with unsupervised feature learning[EB/OL]. [2023-12-19]. http://ufldl.stanford.edu/housenumbers/nips2011_housenumbers.pdf.
40	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE Press, 2016: 770-778.
41	PASZKE A, GROSS S, MASSA F, et al. PyTorch: an imperative style, high-performance deep learning library[EB/OL]. [2023-12-19]. https://papers.nips.cc/paper_files/paper/2019/file/bdbca288fee7f92f2bfa9f7012727740-Paper.pdf.

[1]	孙超, 范之国, 罗茂文, 胡泉. 基于目标场景颜色恢复的偏振图像去雾算法[J]. 计算机工程, 2025, 51(9): 268-279.
[2]	张玉博, 杨帆, 郭亚, 杨文慧. 基于视觉大模型的垃圾分类轻量化算法研究[J]. 计算机工程, 2025, 51(7): 140-151.
[3]	廖丁丁, 刘俊峰, 曾君, 邱晓欢. 一种基于块平均正交权重修正的连续学习算法[J]. 计算机工程, 2025, 51(6): 57-64.
[4]	张政秀, 周淳, 杨萌. 基于实例谱关系的知识蒸馏[J]. 计算机工程, 2025, 51(11): 63-71.
[5]	林烁彬, 蔡捷仪, 方晓城, 张正, 卢光明, 陈炳志. 基于强度相关正则化学习的对抗鲁棒蒸馏方法[J]. 计算机工程, 2025, 51(1): 42-50.
[6]	屠乃威, 焦猛, 阎馨. 复杂环境下输电线路鸟巢目标图像检测模型[J]. 计算机工程, 2024, 50(7): 216-226.
[7]	董良振, 田建艳, 杨胜强, 陈海滨. 基于光照校正和图像融合的零件表面图像增强[J]. 计算机工程, 2024, 50(6): 245-254.
[8]	周昭辰, 方清茂, 吴晓红, 胡平, 何小海. 基于MacBERT与对抗训练的机器阅读理解模型[J]. 计算机工程, 2024, 50(5): 41-50.
[9]	徐宽广, 何东隅, 韩冰, 刘宇佳, 李家栋. 基于机器视觉的手写钢板号图像增强及矫正算法研究与应用[J]. 计算机工程, 2024, 50(4): 350-356.
[10]	董星星, 高继勋, 王晓桐, 李松. 空间方向关系表达与推理模型研究综述[J]. 计算机工程, 2023, 49(9): 1-15.
[11]	艾青松, 张皓喆, 严俊伟. 基于自适应伽马校正的异常驾驶行为检测方法[J]. 计算机工程, 2023, 49(9): 279-286.
[12]	王书朋, 何引弟. 融合特征注意力机制的非均匀光照图像增强算法[J]. 计算机工程, 2023, 49(8): 232-239.
[13]	曹坪, 杨怀志, 薄一军, 尤嘉, 张淳杰, 李丹勇. 面向低质量裂缝图像的多知识蒸馏分类[J]. 计算机工程, 2023, 49(7): 204-213.
[14]	毛亮, 赵林均, 余敦辉, 孙斌. 基于知识蒸馏的企业命名实体识别模型[J]. 计算机工程, 2023, 49(5): 90-96.
[15]	陈何雄, 罗宇薇, 韦云凯, 郭威, 杭菲璐, 何映军, 杨宁. 基于联邦学习的SDN异常流量协同检测技术[J]. 计算机工程, 2023, 49(3): 168-176.

选择文件类型/文献管理软件名称

选择包含的内容