克服神经网络中的灾难性遗忘。
Overcoming catastrophic forgetting in neural networks.
机构信息
DeepMind, London EC4 5TW, United Kingdom;
DeepMind, London EC4 5TW, United Kingdom.
出版信息
Proc Natl Acad Sci U S A. 2017 Mar 28;114(13):3521-3526. doi: 10.1073/pnas.1611835114. Epub 2017 Mar 14.
Abstract
The ability to learn tasks in a sequential fashion is crucial to the development of artificial intelligence. Until now neural networks have not been capable of this and it has been widely thought that catastrophic forgetting is an inevitable feature of connectionist models. We show that it is possible to overcome this limitation and train networks that can maintain expertise on tasks that they have not experienced for a long time. Our approach remembers old tasks by selectively slowing down learning on the weights important for those tasks. We demonstrate our approach is scalable and effective by solving a set of classification tasks based on a hand-written digit dataset and by learning several Atari 2600 games sequentially.
摘要
以序列方式学习任务的能力对人工智能的发展至关重要。到目前为止，神经网络还无法做到这一点，人们普遍认为灾难性遗忘是连接主义模型的一个必然特征。我们表明，克服这一限制并训练能够长时间保持对其未经验证的任务的专业知识的网络是可能的。我们的方法通过选择性地减缓对那些任务重要的权重的学习来记住旧任务。我们通过解决一组基于手写数字数据集的分类任务并顺序学习多个雅达利 2600 游戏来证明我们的方法是可扩展和有效的。
相似文献
1
Overcoming catastrophic forgetting in neural networks.克服神经网络中的灾难性遗忘。
Proc Natl Acad Sci U S A. 2017 Mar 28;114(13):3521-3526. doi: 10.1073/pnas.1611835114. Epub 2017 Mar 14.
2
Memory Recall: A Simple Neural Network Training Framework Against Catastrophic Forgetting.记忆召回：一种针对灾难性遗忘的简单神经网络训练框架。
IEEE Trans Neural Netw Learn Syst. 2022 May;33(5):2010-2022. doi: 10.1109/TNNLS.2021.3099700. Epub 2022 May 2.
3
Overcoming Long-Term Catastrophic Forgetting Through Adversarial Neural Pruning and Synaptic Consolidation.通过对抗性神经修剪和突触巩固来克服长期灾难性遗忘。
IEEE Trans Neural Netw Learn Syst. 2022 Sep;33(9):4243-4256. doi: 10.1109/TNNLS.2021.3056201. Epub 2022 Aug 31.
4
Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization.利用上下文相关门控和突触稳定缓解灾难性遗忘。
Proc Natl Acad Sci U S A. 2018 Oct 30;115(44):E10467-E10475. doi: 10.1073/pnas.1803839115. Epub 2018 Oct 12.
5
Neural modularity helps organisms evolve to learn new skills without forgetting old skills.神经模块化有助于生物体进化，以便在不遗忘旧技能的情况下学习新技能。
PLoS Comput Biol. 2015 Apr 2;11(4):e1004128. doi: 10.1371/journal.pcbi.1004128. eCollection 2015 Apr.
6
Sleep prevents catastrophic forgetting in spiking neural networks by forming a joint synaptic weight representation.睡眠通过形成联合突触权重表示来防止尖峰神经网络中的灾难性遗忘。
PLoS Comput Biol. 2022 Nov 18;18(11):e1010628. doi: 10.1371/journal.pcbi.1010628. eCollection 2022 Nov.
7
Comparing continual task learning in minds and machines.比较心智和机器中的持续任务学习。
Proc Natl Acad Sci U S A. 2018 Oct 30;115(44):E10313-E10322. doi: 10.1073/pnas.1800755115. Epub 2018 Oct 15.
8
Contributions by metaplasticity to solving the Catastrophic Forgetting Problem.介形亚纲对解决灾难性遗忘问题的贡献。
Trends Neurosci. 2022 Sep;45(9):656-666. doi: 10.1016/j.tins.2022.06.002. Epub 2022 Jul 4.
9
Online continual learning with declarative memory.在线使用声明式记忆进行持续学习。
Neural Netw. 2023 Jun;163:146-155. doi: 10.1016/j.neunet.2023.03.025. Epub 2023 Mar 27.
10
Incremental Concept Learning via Online Generative Memory Recall.通过在线生成记忆回忆进行增量概念学习。
IEEE Trans Neural Netw Learn Syst. 2021 Jul;32(7):3206-3216. doi: 10.1109/TNNLS.2020.3010581. Epub 2021 Jul 6.
引用本文的文献
1
ClinicSum: Utilizing Language Models for Generating Clinical Summaries from Patient-Doctor Conversations.临床总结：利用语言模型从医患对话中生成临床总结。
Proc IEEE Int Conf Big Data. 2024 Dec;2024:5050-5059. doi: 10.1109/bigdata62323.2024.10825266. Epub 2025 Jan 16.
2
A robust sampling technique for realistic distribution simulation in federated learning.一种用于联邦学习中逼真分布模拟的强大采样技术。
Int J Comput Assist Radiol Surg. 2025 Sep 2. doi: 10.1007/s11548-025-03504-z.
3
Trajectory Tracking Controller for Quadrotor by Continual Reinforcement Learning in Wind-Disturbed Environment.风力干扰环境下基于持续强化学习的四旋翼轨迹跟踪控制器
Sensors (Basel). 2025 Aug 8;25(16):4895. doi: 10.3390/s25164895.
4
CAREC: Continual Wireless Action Recognition with Expansion-Compression Coordination.CAREC：具有扩展-压缩协调的连续无线动作识别
Sensors (Basel). 2025 Jul 30;25(15):4706. doi: 10.3390/s25154706.
5
Dynamic Network Plasticity and Sample Efficiency in Biological Neural Cultures: A Comparative Study with Deep Reinforcement Learning.生物神经培养中的动态网络可塑性与样本效率：与深度强化学习的比较研究
Cyborg Bionic Syst. 2025 Aug 4;6:0336. doi: 10.34133/cbsystems.0336. eCollection 2025.
6
Dual-Stage Clean-Sample Selection for Incremental Noisy Label Learning.用于增量噪声标签学习的双阶段干净样本选择
Bioengineering (Basel). 2025 Jul 8;12(7):743. doi: 10.3390/bioengineering12070743.
7
Interleaved Replay of Novel and Familiar Memory Traces During Slow-Wave Sleep Prevents Catastrophic Forgetting.慢波睡眠期间新记忆痕迹与熟悉记忆痕迹的交错回放可防止灾难性遗忘。
bioRxiv. 2025 Jun 29:2025.06.25.661579. doi: 10.1101/2025.06.25.661579.
8
Domain-incremental white blood cell classification with privacy-aware continual learning.具有隐私感知持续学习的域增量白细胞分类
Sci Rep. 2025 Jul 15;15(1):25468. doi: 10.1038/s41598-025-08024-z.
9
Fine-Tuning Protein Language Models Unlocks the Potential of Underrepresented Viral Proteomes.微调蛋白质语言模型可释放未充分表征的病毒蛋白质组的潜力。
bioRxiv. 2025 Jun 11:2025.04.17.649224. doi: 10.1101/2025.04.17.649224.
10
BertADP: a fine-tuned protein language model for anti-diabetic peptide prediction.BertADP：用于抗糖尿病肽预测的微调蛋白质语言模型。
BMC Biol. 2025 Jul 15;23(1):210. doi: 10.1186/s12915-025-02312-w.
本文引用的文献
1
What Learning Systems do Intelligent Agents Need? Complementary Learning Systems Theory Updated.智能体需要什么样的学习系统？更新后的补充学习系统理论。
Trends Cogn Sci. 2016 Jul;20(7):512-534. doi: 10.1016/j.tics.2016.05.004.
2
Labelling and optical erasure of synaptic memory traces in the motor cortex.运动皮层中突触记忆痕迹的标记与光学消除
Nature. 2015 Sep 17;525(7569):333-8. doi: 10.1038/nature15257. Epub 2015 Sep 9.
3
Deep learning.深度学习。
Nature. 2015 May 28;521(7553):436-44. doi: 10.1038/nature14539.
4
Branch-specific dendritic Ca(2+) spikes cause persistent synaptic plasticity.特定分支的树突状钙离子峰引发持续性突触可塑性。
Nature. 2015 Apr 9;520(7546):180-5. doi: 10.1038/nature14251. Epub 2015 Mar 30.
5
Human-level control through deep reinforcement learning.通过深度强化学习实现人类水平的控制。
Nature. 2015 Feb 26;518(7540):529-33. doi: 10.1038/nature14236.
6
Sleep promotes branch-specific formation of dendritic spines after learning.睡眠促进学习后树突棘的分支特异性形成。
Science. 2014 Jun 6;344(6188):1173-8. doi: 10.1126/science.1249098.
7
Context-dependent computation by recurrent dynamics in prefrontal cortex.前额叶皮层中依赖上下文的递归动力学计算。
Nature. 2013 Nov 7;503(7474):78-84. doi: 10.1038/nature12742.
8
Cognitive control over learning: creating, clustering, and generalizing task-set structure.认知控制学习：创建、聚类和泛化任务集结构。
Psychol Rev. 2013 Jan;120(1):190-229. doi: 10.1037/a0030852.
9
Synapses with short-term plasticity are optimal estimators of presynaptic membrane potentials.具有短期可塑性的突触是最佳的突触前膜电位估计算子。
Nat Neurosci. 2010 Oct;13(10):1271-5. doi: 10.1038/nn.2640. Epub 2010 Sep 19.
10
Stably maintained dendritic spines are associated with lifelong memories.稳定维持的树突棘与终身记忆有关。
Nature. 2009 Dec 17;462(7275):920-4. doi: 10.1038/nature08577. Epub 2009 Nov 29.
Zotero 插件