终身学习的机器:持续学习的根本问题与最有希望的方向

参考文献(全部为公开文献)

正文以 [S编号] 引用下列条目。所有条目以 arXiv 编号或 DOI 为准;会议/期刊名称仅在经直接核验后标注。

A. 问题定义与综述

[S1] McCloskey, M. & Cohen, N. J. (1989). Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation, Vol. 24, 109–165. — 灾难性遗忘的首次系统报告。

[S2] Ratcliff, R. (1990). Connectionist models of recognition memory: Constraints imposed by learning and forgetting functions. Psychological Review, 97(2), 285–308. — 早期对连接主义遗忘的定量刻画。

[S3] French, R. M. (1999). Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4), 128–135. — 经典综述;提出分布式表示的重叠是遗忘根源。

[S4] Grossberg, S. (1980). How does a brain build a cognitive code? Psychological Review, 87(1), 1–51. — "稳定性–可塑性两难"(stability–plasticity dilemma)概念来源。

[S5] Carpenter, G. A. & Grossberg, S. (1987). A massively parallel architecture for a self-organizing neural pattern recognition machine. Computer Vision, Graphics, and Image Processing, 37, 54–115. — 自适应共振理论(ART),最早直面稳定性–可塑性的架构。

[S6] Ring, M. B. (1994). Continual Learning in Reinforcement Environments. PhD thesis, University of Texas at Austin. — "continual learning" 一词在强化学习语境中的奠基性论述。

[S7] Thrun, S. & Mitchell, T. M. (1995). Lifelong robot learning. Robotics and Autonomous Systems, 15, 25–46. — 终身学习(lifelong learning)问题的早期系统提法。

[S8] Chen, Z. & Liu, B. (2018). Lifelong Machine Learning (2nd ed.). Morgan & Claypool. — 终身机器学习专著。

[S9] Mitchell, T. et al. (2018). Never-Ending Learning. Communications of the ACM, 61(5), 103–115. — NELL:运行多年的持续学习系统实例。

[S10] Parisi, G. I., Kemker, R., Part, J. L., Kanan, C. & Wermter, S. (2019). Continual lifelong learning with neural networks: A review. Neural Networks, 113, 54–71. — 神经网络持续学习综述。

[S11] Hadsell, R., Rao, D., Rusu, A. A. & Pascanu, R. (2020). Embracing Change: Continual Learning in Deep Neural Networks. Trends in Cognitive Sciences, 24(12), 1028–1040. — DeepMind 视角的持续学习综述。

[S12] van de Ven, G. M. & Tolias, A. S. (2019). Three scenarios for continual learning. arXiv:1904.07734. https://arxiv.org/abs/1904.07734 — 任务增量/域增量/类增量三种场景的标准划分。

[S13] van de Ven, G. M., Tuytelaars, T. & Tolias, A. S. (2022). Three types of incremental learning. Nature Machine Intelligence, 4, 1185–1197. — [S12] 的期刊扩展版;系统比较三种场景下各类方法的成败。

[S14] Wang, L., Zhang, X., Su, H. & Zhu, J. (2024). A Comprehensive Survey of Continual Learning: Theory, Method and Application. IEEE TPAMI. arXiv:2302.00487. https://arxiv.org/abs/2302.00487 — 目前覆盖面最广的持续学习综述之一。

[S15] Lopez-Paz, D. & Ranzato, M. (2017). Gradient Episodic Memory for Continual Learning. NeurIPS 2017. arXiv:1706.08840. https://arxiv.org/abs/1706.08840 — GEM;正式定义了前向/后向迁移(FWT/BWT)指标。

[S16] Farquhar, S. & Gal, Y. (2018). Towards Robust Evaluations of Continual Learning. arXiv:1805.09733. https://arxiv.org/abs/1805.09733 — 对标准评测协议的系统批评。

[S17] Prabhu, A., Torr, P. H. S. & Dokania, P. K. (2020). GDumb: A Simple Approach that Questions Our Progress in Continual Learning. ECCV 2020, LNCS 12347, 524–540. — 贪心存样本+从头重训的朴素基线胜过大量复杂方法,对领域进展提出质疑。

[S18] Knoblauch, J., Husain, H. & Diethe, T. (2020). Optimal Continual Learning has Perfect Memory and is NP-hard. ICML 2020, PMLR 119, 5327–5337. arXiv:2006.05188. https://arxiv.org/abs/2006.05188 — 理论结果:最优持续学习一般需要完美记忆且为 NP 难。

[S19] Aljundi, R., Kelchtermans, K. & Tuytelaars, T. (2019). Task-Free Continual Learning. CVPR 2019. arXiv:1812.03596. https://arxiv.org/abs/1812.03596 — 无任务边界的持续学习设定。

[S20] Silver, D. & Sutton, R. S. (2025). Welcome to the Era of Experience. Preprint of a chapter in "Designing an Intelligence" (MIT Press). https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf — 主张智能体应栖居于终身经验流并从中持续学习。

[S21] Sutton, R. S. (2019). The Bitter Lesson. 公开博客文章. http://www.incompleteideas.net/IncIdeas/BitterLesson.html — 通用方法+算力胜过人工先验的著名论断。

[S22] Kudithipudi, D. et al. (2022). Biological underpinnings for lifelong learning machines. Nature Machine Intelligence, 4, 196–210. — 生物终身学习机制面向机器的系统盘点。

B. 可塑性丧失与遗忘机制

[S23] Dohare, S., Hernandez-Garcia, J. F., Lan, Q., Rahman, P., Mahmood, A. R. & Sutton, R. S. (2024). Loss of plasticity in deep continual learning. Nature, 632, 768–774. DOI: 10.1038/s41586-024-07711-7. https://www.nature.com/articles/s41586-024-07711-7 — 标准深度学习在持续设定下逐渐丧失可塑性;提出持续反向传播(continual backprop)。

[S24] Goodfellow, I. J., Mirza, M., Xiao, D., Courville, A. & Bengio, Y. (2013). An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks. arXiv:1312.6211. https://arxiv.org/abs/1312.6211 — 深度学习时代对灾难性遗忘的早期实证研究。

[S25] Ash, J. T. & Adams, R. P. (2020). On Warm-Starting Neural Network Training. NeurIPS 2020. arXiv:1910.08475. https://arxiv.org/abs/1910.08475 — 温启动损害泛化;提出 shrink-and-perturb。

[S26] Nikishin, E., Schwarzer, M., D'Oro, P., Bacon, P.-L. & Courville, A. (2022). The Primacy Bias in Deep Reinforcement Learning. ICML 2022. arXiv:2205.07802. https://arxiv.org/abs/2205.07802 — 深度 RL 中的"首因偏差"与周期性重置。

[S27] Sokar, G., Agarwal, R., Castro, P. S. & Evci, U. (2023). The Dormant Neuron Phenomenon in Deep Reinforcement Learning. ICML 2023. arXiv:2302.12902. https://arxiv.org/abs/2302.12902 — 休眠神经元随训练累积,削弱表达能力。

[S28] Lyle, C., Zheng, Z., Nikishin, E., Avila Pires, B., Pascanu, R. & Dabney, W. (2023). Understanding Plasticity in Neural Networks. ICML 2023. arXiv:2303.01486. https://arxiv.org/abs/2303.01486 — 可塑性丧失的机理分析(曲率、参数增长等)。

[S29] Ramasesh, V. V., Dyer, E. & Raghu, M. (2021). Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics. ICLR 2021. arXiv:2007.07400. https://arxiv.org/abs/2007.07400 — 遗忘主要由深层表示的漂移驱动。

[S30] Ramasesh, V. V., Lewkowycz, A. & Dyer, E. (2022). Effect of scale on catastrophic forgetting in neural networks. ICLR 2022. https://openreview.net/forum?id=GhVS8_yPeEa — 模型规模越大、预训练越充分,遗忘越轻。

[S31] Kumar, S., Marklund, H. & Van Roy, B. (2023). Maintaining Plasticity in Continual Learning via Regenerative Regularization. arXiv:2308.11958. https://arxiv.org/abs/2308.11958 — L2 Init:向初始参数正则化以保持可塑性。

C. 损失几何与"就近学习"

[S32] Garipov, T., Izmailov, P., Podoprikhin, D., Vetrov, D. & Wilson, A. G. (2018). Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs. NeurIPS 2018. arXiv:1802.10026. https://arxiv.org/abs/1802.10026 — 独立训练的极小点由低损失曲线相连(模式连通性)。

[S33] Draxler, F., Veschgini, K., Salmhofer, M. & Hamprecht, F. A. (2018). Essentially No Barriers in Neural Network Energy Landscape. ICML 2018. arXiv:1803.00885. https://arxiv.org/abs/1803.00885 — 与 [S32] 同期独立发现:极小点之间基本无能量壁垒。

[S34] Frankle, J., Dziugaite, G. K., Roy, D. M. & Carbin, M. (2020). Linear Mode Connectivity and the Lottery Ticket Hypothesis. ICML 2020. arXiv:1912.05671. https://arxiv.org/abs/1912.05671 — 线性模式连通性(LMC)与训练早期的"稳定化"。

[S35] Entezari, R., Sedghi, H., Saukh, O. & Neyshabur, B. (2022). The Role of Permutation Invariance in Linear Mode Connectivity of Neural Networks. ICLR 2022. arXiv:2110.06296. https://arxiv.org/abs/2110.06296 — 猜想:模除置换对称后,SGD 解基本落在同一线性连通盆地。

[S36] Ainsworth, S. K., Hayase, J. & Srinivasa, S. (2023). Git Re-Basin: Merging Models modulo Permutation Symmetries. ICLR 2023. arXiv:2209.04836. https://arxiv.org/abs/2209.04836 — 通过置换对齐合并独立训练的模型。

[S37] Jacot, A., Gabriel, F. & Hongler, C. (2018). Neural Tangent Kernel: Convergence and Generalization in Neural Networks. NeurIPS 2018. arXiv:1806.07572. https://arxiv.org/abs/1806.07572 — NTK:无限宽极限下网络训练近似线性化。

[S38] Chizat, L., Oyallon, E. & Bach, F. (2019). On Lazy Training in Differentiable Programming. NeurIPS 2019. arXiv:1812.07956. https://arxiv.org/abs/1812.07956 — "惰性训练"机制及其与特征学习的对立。

[S39] Mirzadeh, S. I., Farajtabar, M., Pascanu, R. & Ghasemzadeh, H. (2020). Understanding the Role of Training Regimes in Continual Learning. NeurIPS 2020. arXiv:2006.06958. https://arxiv.org/abs/2006.06958 — 平坦极小点与遗忘的关系;训练制度(学习率、批大小)影响遗忘。

[S40] Mirzadeh, S. I., Farajtabar, M., Gorur, D., Pascanu, R. & Ghasemzadeh, H. (2021). Linear Mode Connectivity in Multitask and Continual Learning. ICLR 2021. arXiv:2010.04495. https://arxiv.org/abs/2010.04495 — 同初始化下,多任务解与持续学习解线性连通;提出 MC-SGD。

[S41] Kirkpatrick, J. et al. (2017). Overcoming catastrophic forgetting in neural networks. PNAS, 114(13), 3521–3526. DOI: 10.1073/pnas.1611835114. arXiv:1612.00796 — EWC:以 Fisher 信息加权的二次惩罚保护旧任务重要参数。

[S42] Zenke, F., Poole, B. & Ganguli, S. (2017). Continual Learning Through Synaptic Intelligence. ICML 2017. arXiv:1703.04200. https://arxiv.org/abs/1703.04200 — SI:在线累积参数重要性。

[S43] Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M. & Tuytelaars, T. (2018). Memory Aware Synapses: Learning what (not) to forget. ECCV 2018. arXiv:1711.09601. https://arxiv.org/abs/1711.09601 — MAS:无监督估计参数重要性。

[S44] Farajtabar, M., Azizan, N., Mott, A. & Li, A. (2020). Orthogonal Gradient Descent for Continual Learning. AISTATS 2020. arXiv:1910.07104. https://arxiv.org/abs/1910.07104 — OGD:把新任务梯度投影到旧任务梯度的正交补。

[S45] Zeng, G., Chen, Y., Cui, B. & Yu, S. (2019). Continual learning of context-dependent processing in neural networks. Nature Machine Intelligence, 1, 364–372. — OWM:正交权重修改+情境依赖处理模块。

[S46] Saha, G., Garg, I. & Roy, K. (2021). Gradient Projection Memory for Continual Learning. ICLR 2021. arXiv:2103.09762. https://arxiv.org/abs/2103.09762 — GPM:在旧任务核心梯度子空间的正交方向学习。

[S47] Hu, E. J. et al. (2022). LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022. arXiv:2106.09685. https://arxiv.org/abs/2106.09685 — 低秩增量更新:冻结主干、只学小的低秩修正。

[S48] Ilharco, G. et al. (2023). Editing Models with Task Arithmetic. ICLR 2023. arXiv:2212.04089. https://arxiv.org/abs/2212.04089 — 任务向量可加减:权重空间中的知识算术。

[S49] Wortsman, M. et al. (2022). Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. ICML 2022. arXiv:2203.05482. https://arxiv.org/abs/2203.05482 — 同一预训练模型的多个微调解可以直接平均。

[S50] Wortsman, M. et al. (2022). Robust fine-tuning of zero-shot models. CVPR 2022. arXiv:2109.01903. https://arxiv.org/abs/2109.01903 — WiSE-FT:微调权重与原权重线性插值兼得新旧性能。

D. 系统路线:重放、双系统、记忆、LLM 与智能体

[S51] Robins, A. (1995). Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2), 123–146. — 重放与伪重放(用自生成样本复习)的奠基工作。

[S52] Rebuffi, S.-A., Kolesnikov, A., Sperl, G. & Lampert, C. H. (2017). iCaRL: Incremental Classifier and Representation Learning. CVPR 2017. arXiv:1611.07725. https://arxiv.org/abs/1611.07725 — 类增量学习的代表性重放方法。

[S53] Shin, H., Lee, J. K., Kim, J. & Kim, J. (2017). Continual Learning with Deep Generative Replay. NeurIPS 2017. arXiv:1705.08690. https://arxiv.org/abs/1705.08690 — 生成式重放:用生成模型合成旧数据。

[S54] Chaudhry, A., Ranzato, M., Rohrbach, M. & Elhoseiny, M. (2019). Efficient Lifelong Learning with A-GEM. ICLR 2019. arXiv:1812.00420. https://arxiv.org/abs/1812.00420 — GEM 的高效近似。

[S55] Buzzega, P., Boschini, M., Porrello, A., Abati, D. & Calderara, S. (2020). Dark Experience for General Continual Learning: a Strong, Simple Baseline. NeurIPS 2020. arXiv:2004.07211. https://arxiv.org/abs/2004.07211 — DER:重放 logits 的强基线。

[S56] Rusu, A. A. et al. (2016). Progressive Neural Networks. arXiv:1606.04671. https://arxiv.org/abs/1606.04671 — 每个任务新增一列网络,结构上杜绝遗忘。

[S57] Schwarz, J. et al. (2018). Progress & Compress: A scalable framework for continual learning. ICML 2018. arXiv:1805.06370. https://arxiv.org/abs/1805.06370 — 快速学习列+知识蒸馏固化到慢速列。

[S58] Mallya, A. & Lazebnik, S. (2018). PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning. CVPR 2018. arXiv:1711.05769. https://arxiv.org/abs/1711.05769 — 剪枝腾挪容量、按任务分配子网络。

[S59] Serra, J., Suris, D., Miron, M. & Karatzoglou, A. (2018). Overcoming Catastrophic Forgetting with Hard Attention to the Task. ICML 2018. arXiv:1801.01423. https://arxiv.org/abs/1801.01423 — HAT:学习任务级硬注意力掩码。

[S60] Masse, N. Y., Grant, G. D. & Freedman, D. J. (2018). Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization. PNAS, 115(44), E10467–E10475. DOI: 10.1073/pnas.1803839115. arXiv:1802.01569 — XdG:稀疏情境门控+突触稳固,在数百个任务上保持高性能。

[S61] McClelland, J. L., McNaughton, B. L. & O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex. Psychological Review, 102(3), 419–457. — 互补学习系统(CLS)理论。

[S62] Kumaran, D., Hassabis, D. & McClelland, J. L. (2016). What Learning Systems do Intelligent Agents Need? Complementary Learning Systems Theory Updated. Trends in Cognitive Sciences, 20(7), 512–534. — CLS 理论的更新版,面向智能体设计。

[S63] Graves, A. et al. (2016). Hybrid computing using a neural network with dynamic external memory. Nature, 538, 471–476. — 可微神经计算机(DNC):可读写外部记忆。

[S64] Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020. arXiv:2005.11401. https://arxiv.org/abs/2005.11401 — RAG:检索增强生成。

[S65] Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L. & Lewis, M. (2020). Generalization through Memorization: Nearest Neighbor Language Models. ICLR 2020. arXiv:1911.00172. https://arxiv.org/abs/1911.00172 — kNN-LM:显式记忆库改进语言模型。

[S66] Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A. A. & Hardt, M. (2020). Test-Time Training with Self-Supervision for Generalization under Distribution Shifts. ICML 2020. arXiv:1909.13231. https://arxiv.org/abs/1909.13231 — 测试时训练(TTT)。

[S67] Sun, Y. et al. (2024). Learning to (Learn at Test Time): RNNs with Expressive Hidden States. arXiv:2407.04620. https://arxiv.org/abs/2407.04620 — TTT 层:把隐状态本身变成一个在推理中被训练的模型。

[S68] Behrouz, A., Zhong, P. & Mirrokni, V. (2024). Titans: Learning to Memorize at Test Time. arXiv:2501.00663. https://arxiv.org/abs/2501.00663 — 神经长期记忆模块在推理时按"惊讶度"做梯度写入,带自适应遗忘。

[S69] Brown, T. B. et al. (2020). Language Models are Few-Shot Learners. NeurIPS 2020. arXiv:2005.14165. https://arxiv.org/abs/2005.14165 — GPT-3;上下文内学习(ICL)现象的确立。

[S70] von Oswald, J. et al. (2023). Transformers Learn In-Context by Gradient Descent. ICML 2023. arXiv:2212.07677. https://arxiv.org/abs/2212.07677 — ICL 可实现为前向传播中的隐式梯度下降。

[S71] Akyürek, E., Schuurmans, D., Andreas, J., Ma, T. & Zhou, D. (2023). What learning algorithm is in-context learning? Investigations with linear models. ICLR 2023. arXiv:2211.15661. https://arxiv.org/abs/2211.15661 — ICL 在线性模型上隐式实现标准学习算法。

[S72] Gururangan, S. et al. (2020). Don't Stop Pretraining: Adapt Language Models to Domains and Tasks. ACL 2020. arXiv:2004.10964. https://arxiv.org/abs/2004.10964 — 领域/任务自适应持续预训练(DAPT/TAPT)。

[S73] Ke, Z., Shao, Y., Lin, H., Konishi, T., Kim, G. & Liu, B. (2023). Continual Pre-training of Language Models. ICLR 2023. arXiv:2302.03241. https://arxiv.org/abs/2302.03241 — 语言模型持续预训练的软掩码方法。

[S74] Ibrahim, A., Thérien, B., Gupta, K. et al. (2024). Simple and Scalable Strategies to Continually Pre-train Large Language Models. arXiv:2403.08763. https://arxiv.org/abs/2403.08763 — 学习率再升温+小比例重放即可接近重新训练的效果。

[S75] Scialom, T., Chakrabarty, T. & Muresan, S. (2022). Fine-tuned Language Models are Continual Learners. EMNLP 2022. arXiv:2205.12393. https://arxiv.org/abs/2205.12393 — 指令微调模型在少量重放下的持续学习能力。

[S76] Luo, Y., Yang, Z., Meng, F., Li, Y., Zhou, J. & Zhang, Y. (2023). An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning. arXiv:2308.08747. https://arxiv.org/abs/2308.08747 — 1B–7B 级 LLM 持续微调中的遗忘实证。

[S77] Meng, K., Bau, D., Andonian, A. & Belinkov, Y. (2022). Locating and Editing Factual Associations in GPT. NeurIPS 2022. arXiv:2202.05262. https://arxiv.org/abs/2202.05262 — ROME:定位并编辑事实关联。

[S78] Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y. & Bau, D. (2023). Mass-Editing Memory in a Transformer. ICLR 2023. arXiv:2210.07229. https://arxiv.org/abs/2210.07229 — MEMIT:批量知识编辑及其规模极限。

[S79] Wang, G. et al. (2023). Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv:2305.16291. https://arxiv.org/abs/2305.16291 — 不改权重、以代码技能库积累能力的开放世界智能体。

[S80] Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K. & Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023. arXiv:2303.11366. https://arxiv.org/abs/2303.11366 — 用语言化自我反思(存于情景记忆)代替权重更新。

[S81] Hinton, G. E., Dayan, P., Frey, B. J. & Neal, R. M. (1995). The "wake-sleep" algorithm for unsupervised neural networks. Science, 268, 1158–1161. — 醒–睡两阶段学习算法。

[S82] Tadros, T., Krishnan, G. P., Ramyaa, R. & Bazhenov, M. (2022). Sleep-like unsupervised replay reduces catastrophic forgetting in artificial neural networks. Nature Communications, 13, 7742. DOI: 10.1038/s41467-022-34938-7 — 睡眠样自发重放(局部 Hebbian 规则+噪声输入)恢复被遗忘的旧任务。

E. 神经科学

[S83] Yang, G., Pan, F. & Gan, W.-B. (2009). Stably maintained dendritic spines are associated with lifelong memories. Nature, 462, 920–924. — 树突棘可稳定保持一生,承载长期记忆。

[S84] Yang, G. et al. (2014). Sleep promotes branch-specific formation of dendritic spines after learning. Science, 344, 1173–1178. — 睡眠促进学习后分支特异的新棘形成。

[S85] Cichon, J. & Gan, W.-B. (2015). Branch-specific dendritic Ca²⁺ spikes cause persistent synaptic plasticity. Nature, 520, 180–185. — 不同任务的可塑性被分配到不同树突分支,减少干扰。

[S86] Wilson, M. A. & McNaughton, B. L. (1994). Reactivation of hippocampal ensemble memories during sleep. Science, 265, 676–679. — 海马在睡眠中重放清醒期的神经活动序列。

[S87] McGaugh, J. L. (2000). Memory—a Century of Consolidation. Science, 287, 248–251. — 记忆固化研究百年综述。

[S88] Tononi, G. & Cirelli, C. (2014). Sleep and the Price of Plasticity. Neuron, 81(1), 12–34. — 突触稳态假说(SHY):睡眠期间全局突触降标。

[S89] Fusi, S., Drew, P. J. & Abbott, L. F. (2005). Cascade models of synaptically stored memories. Neuron, 45(4), 599–611. — 多状态级联突触:用内部隐状态权衡记忆强度与寿命。

[S90] Benna, M. K. & Fusi, S. (2016). Computational principles of synaptic memory consolidation. Nature Neuroscience, 19, 1697–1706. — 多时间尺度变量链的突触模型,遗忘呈幂律衰减。

[S91] Abraham, W. C. & Bear, M. F. (1996). Metaplasticity: the plasticity of synaptic plasticity. Trends in Neurosciences, 19(4), 126–130. — 元可塑性:突触改变自身可塑性的能力。

[S92] Olshausen, B. A. & Field, D. J. (1996). Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 381, 607–609. — 稀疏编码的奠基工作。

[S93] Eriksson, P. S. et al. (1998). Neurogenesis in the adult human hippocampus. Nature Medicine, 4, 1313–1317. — 成人海马神经发生的经典证据。

[S94] Sorrells, S. F. et al. (2018). Human hippocampal neurogenesis drops sharply in children to undetectable levels in adults. Nature, 555, 377–381. DOI: 10.1038/nature25975 — 成人海马神经发生未检出(与 [S95] 直接冲突)。

[S95] Boldrini, M. et al. (2018). Human Hippocampal Neurogenesis Persists throughout Aging. Cell Stem Cell, 22, 589–599. DOI: 10.1016/j.stem.2018.03.015 — 成人海马神经发生持续到 79 岁(与 [S94] 直接冲突)。

[S96] Flesch, T., Balaguer, J., Dekker, R., Nili, H. & Summerfield, C. (2018). Comparing continual task learning in minds and machines. PNAS, 115(44), E10313–E10322. — 人类在分块(blocked)训练下学得更好,与标准神经网络相反。

[S97] Hassabis, D., Kumaran, D., Summerfield, C. & Botvinick, M. (2017). Neuroscience-Inspired Artificial Intelligence. Neuron, 95, 245–258. — 神经科学启发 AI 的路线图。

[S98] Richards, B. A. et al. (2019). A deep learning framework for neuroscience. Nature Neuroscience, 22, 1761–1770. — 用深度学习三要素(目标、结构、规则)组织神经科学。

[S99] Aston-Jones, G. & Cohen, J. D. (2005). An integrative theory of locus coeruleus-norepinephrine function: adaptive gain and optimal performance. Annual Review of Neuroscience, 28, 403–450. — 去甲肾上腺素系统对学习/探索增益的调控。

F. 补充文献(2026-09 更新)

F.1 正则化与"就近"方法

[S100] Li, Z. & Hoiem, D. (2016). Learning without Forgetting. arXiv:1606.09282. https://arxiv.org/abs/1606.09282 — LwF:蒸馏式函数空间正则,让新模型在新数据上模仿旧模型的输出。

[S101] Li, X., Grandvalet, Y. & Davoine, F. (2018). Explicit Inductive Bias for Transfer Learning with Convolutional Networks. arXiv:1802.01483. https://arxiv.org/abs/1802.01483 — L2-SP:向预训练起点(而非原点)做 L2 正则,是"拴在旧参数附近"最直接的经典形式。

[S102] Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. (2018). Experience Replay for Continual Learning. arXiv:1811.11682. https://arxiv.org/abs/1811.11682 — 强化学习设定下经验重放的系统研究。

[S103] van de Ven, G. M., Siegelmann, H. T. & Tolias, A. S. (2020). Brain-inspired replay for continual learning with artificial neural networks. Nature Communications, 11. DOI: 10.1038/s41467-020-17866-2 — 类脑生成式重放:重放内部表示而非原始像素,类增量场景下大幅改进。

F.2 几何与理论

[S104] Wang, Z. et al. (2021). Learning to Prompt for Continual Learning (L2P). arXiv:2112.08654. https://arxiv.org/abs/2112.08654 — 冻结预训练主干,只学一组可检索的提示向量。

[S105] Wang, Z. et al. (2022). DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning. arXiv:2204.04799. https://arxiv.org/abs/2204.04799 — 无重放的双提示持续学习。

[S106] Mirzadeh, S. I. et al. (2021). Wide Neural Networks Forget Less Catastrophically. arXiv:2110.11526. https://arxiv.org/abs/2110.11526 — 网络越宽遗忘越少;系统研究宽度/深度对遗忘的影响。

[S107] Doan, T. et al. (2020). A Theoretical Analysis of Catastrophic Forgetting through the NTK Overlap Matrix. arXiv:2010.04003. https://arxiv.org/abs/2010.04003 — 用 NTK 重叠矩阵定量刻画任务间干扰与遗忘。

[S108] Bennani, M. A., Doan, T. & Sugiyama, M. (2020). Generalisation Guarantees for Continual Learning with Orthogonal Gradient Descent. arXiv:2006.11942. https://arxiv.org/abs/2006.11942 — NTK 区内 OGD 的泛化理论保证。

[S109] Evron, I. et al. (2022). How catastrophic can catastrophic forgetting be in linear regression? arXiv:2205.09588. https://arxiv.org/abs/2205.09588 — 线性回归中遗忘的精确理论刻画。

[S110] Fort, S. & Jastrzebski, S. (2019). Large Scale Structure of Neural Network Loss Landscapes. arXiv:1906.04724. https://arxiv.org/abs/1906.04724 — 高维损失地形的大尺度结构模型。

[S111] Li, C., Farkhoor, H., Liu, R. & Yosinski, J. (2018). Measuring the Intrinsic Dimension of Objective Landscapes. arXiv:1804.08838. https://arxiv.org/abs/1804.08838 — 许多任务的"内在维度"远低于参数总数。

[S112] Lubana, E. S. et al. (2022). Mechanistic Mode Connectivity. arXiv:2211.08422. https://arxiv.org/abs/2211.08422 — 损失连通的两个解可能采用完全不同的内部机制;损失接近≠机制接近。

F.3 LLM 时代的持续学习

[S113] Biderman, D. et al. (2024). LoRA Learns Less and Forgets Less. arXiv:2405.09673. https://arxiv.org/abs/2405.09673 — 大规模实证:低秩更新学到的更少、忘掉的也更少——"就近学习"权衡在 LLM 尺度上的直接证据。

[S114] Wang, X. et al. (2023). Orthogonal Subspace Learning for Language Model Continual Learning (O-LoRA). arXiv:2310.14152. https://arxiv.org/abs/2310.14152 — 在相互正交的低秩子空间中顺序学习任务。

[S115] Yadav, P. et al. (2023). TIES-Merging: Resolving Interference When Merging Models. arXiv:2306.01708. https://arxiv.org/abs/2306.01708 — 定位并消解模型合并中的参数干扰(符号冲突、冗余)。

[S116] Matena, M. & Raffel, C. (2021). Merging Models with Fisher-Weighted Averaging. arXiv:2111.09832. https://arxiv.org/abs/2111.09832 — 以 Fisher 信息加权的模型合并。

[S117] Yoo, H. et al. (2025). Continual Low-Rank Adapters for LLM-based Generative Recommender Systems (PESO). ICLR 2026(会议页面已核验)。arXiv:2510.25093. https://arxiv.org/abs/2510.25093 — 单个持续演化的 LoRA,以近端项锚定其最近的冻结状态;并明确论证在推荐场景中过时知识应当被遗忘。

[S118] Qiao, F. & Mahdavi, M. (2025). Merge before Forget: A Single LoRA Continual Learning via Continual Merging (SLAO). ICLR 2026(论文集页面已核验)。arXiv:2512.23017. https://arxiv.org/abs/2512.23017 — 正交初始化+时间感知缩放,把任务更新持续合并进单个 LoRA,内存不随任务数增长。

[S119] Gupta, K. et al. (2023). Continual Pre-Training of Large Language Models: How to (re)warm your model? arXiv:2308.04014. https://arxiv.org/abs/2308.04014 — 学习率重新升温对持续预训练的关键作用。

[S120] Gupta, A., Rao, A. & Anumanchipalli, G. (2024). Model Editing at Scale leads to Gradual and Catastrophic Forgetting. arXiv:2401.07453. https://arxiv.org/abs/2401.07453 — 连续编辑呈两阶段退化:先渐进遗忘,后被单次"致瘫编辑"突然击溃。

[S121] Mitchell, E. et al. (2021). Fast Model Editing at Scale (MEND). arXiv:2110.11309. https://arxiv.org/abs/2110.11309 — 超网络式快速模型编辑。

[S122] Mai, Z. et al. (2021). Online Continual Learning in Image Classification: An Empirical Survey. arXiv:2101.10423. https://arxiv.org/abs/2101.10423 — 在线(单遍、无边界)持续学习的实证综述。

[S123] De Lange, M. et al. (2019). A continual learning survey: Defying forgetting in classification tasks. arXiv:1909.08383. https://arxiv.org/abs/1909.08383 — 分类任务持续学习的系统综述。

[S124] Prabhu, A. et al. (2023). Computationally Budgeted Continual Learning: What Does Matter? arXiv:2303.11165. https://arxiv.org/abs/2303.11165 — 固定算力预算下重新评估:许多已发表的增益在预算对齐比较中消失,简单方法常胜出。

[S125] Wu, T. et al. (2024). Continual Learning for Large Language Models: A Survey. arXiv:2402.01364. https://arxiv.org/abs/2402.01364 — LLM 持续学习综述。

[S126] Shi, H. et al. (2024). Continual Learning of Large Language Models: A Comprehensive Survey. arXiv:2404.16789. https://arxiv.org/abs/2404.16789 — LLM 持续学习综述(另一组)。

F.4 相关但间接

[S127] Li, Jinhao et al. (2026). Signal-Adaptive Trust Regions for Gradient-Free Optimization of Recurrent Spiking Neural Networks. arXiv:2601.21572(公开预印本). https://arxiv.org/abs/2601.21572 — 强化学习/优化方向的工作,与持续学习仅间接相关:体现"把每步更新限制在分布空间的一个信赖域内"的一般原则。本文仅按预印本引用。

核实说明:标注 DOI 或 arXiv 号的条目均可直接访问;[S17][S18][S20][S23][S31][S40][S60][S68][S82][S94][S95] 的书目信息在成文时经在线逐条核对。S23 的 arXiv 预印本版本为 arXiv:2306.13812(预印本标题 "Maintaining Plasticity in Deep Continual Learning",发表版标题不同)。F 节(S100–S127)条目的标题/作者/年份经 arXiv 摘要页或 Crossref 核验;其中会议名称仅在 S103与 [S117][S118](ICLR 2026 官方页面)处经直接核验后标注,其余条目不标注会议名称,以 arXiv 编号为准。若发现任何条目有误,请以原文为准并告知我们更正。