关于持续学习的24、25年顶会论文
1、Train-Attention: Meta-Learning Where to Focus in Continual Knowledge Learning
动态预测token的重要性,槽位?有点复杂
2、Bridge Past and Future: Overcoming Information Asymmetry in Incremental Object Detection
过去现在未来,用背景连接未来?蒸馏,eccv2024【思路偏图片,背景,现在不用】
3、Adapt without Forgetting: Distill Proximity from Dual Teachers in Vision-Language Models
数据未知场景的蒸馏,图结构表示,引入两个教师模型(原始CLIP模型和前一个任务的模型),通过蒸馏技术将这些教师模型的知识传递给学生模型
4、Online Continuous Generalized Category Discovery
通过能量引导发现将数据分为已知、见过和未见类别,未见数据未标记,方差集特征增强改善聚类,参数调整基于能量的对比损失
5、Improving Plasticity in Online Continual Learning via Collaborative Learning (CVPR2024)
未训练网络 θ2 的预测对网络 θ1 进行蒸馏,协同持续学习,引入两个架构相同且优化器设置相同的持续学习者,以一种互教的方式同时从数据流中学习。两个网络不仅接受真实标签的监督,还接受彼此预测的监督。蒸馏生成,从更难的样本到更容易的样本进行logit分布的蒸馏
6、Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning (CVPR2024)
利用生成对抗样本来模拟旧类数据,从而估计嵌入空间的漂移
7、Mask and Compress: Efficient Skeleton-based Action Recognition in Continual Learning
在der上改的,有对比。均匀采样和线性插值压缩骨架序列,随机掩码输入数据并进行重建,在每个任务结束后,冻结编码器的参数,只优化分类器的参数,解决训练和测试时输入条件不一致的问题
8、Improvements of Dark Experience Replay and Reservoir Sampling towards Better Balance between Consolidation and Plasticity
数据增强+分类任务,对der的改进,没有代码
引入 q-对数函数(fq(n))作为计数器,调整数据被接受的概率,当 q 较小时,RS 缓冲区更倾向于巩固旧数据;引入多个 RS 缓冲区,每个缓冲区都有不同的 q 值和时间尺度,浅层缓冲区(如第一个缓冲区)更倾向于可塑性,深层缓冲区(如最后一个缓冲区)更倾向于巩固;根据数据的重要性(通过重放优先级计算)决定是否省略该数据
9、Loss of plasticity in deep continual learning
Nature2025通过算法continually inject diversity into the network, small fraction of less-used units are continually and randomly reinitialized l2正则化,网络神经元方式
10、LOIRE: LIFELONG LEARNING ON INCREMENTAL DATA VIA PRE-TRAINED LANGUAGE MODEL GROWTH EFFICIENTLY
2025iclr,网络框架的隐藏层,生成优化序列,教师学生蒸馏,但是不知道实验模型和数据,是在预训练模型,不是针对任务
11、Task Confusion and Catastrophic Forgetting in Class-Incremental Learning: A Mathematical Framework for Discriminative and Generative Modelings
任务的混淆,有正则、蒸馏、bias相关性、生成replay、生成的clser
12、Disentangling and mitigating the impact of task similarity for continual learning
任务之间的相似性,2024,似乎偏理论
13、Multi-layer Rehearsal Feature Augmentation for Class-Incremental Learning
旧样本的复现
14、Coherent Temporal Synthesis for Incremental Action Segmentation (CVPR2024)
15、Text-Enhanced Data-free Approach for Federated Class-Incremental Learning (CVPR2024)
16、SDDGR: Stable Diffusion-based Deep Generative Replay for Class Incremental Object Detection (CVPR2024)
17、Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR2024)
18、Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster Memory (CVPR2024)
19、eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI2024)
20、Plasticity-Optimized Complementary Networks for Unsupervised Continual (WACV2024)
1、Improving Plasticity in Online Continual Learning via Collaborative Learning
2024cvpr,稳定性可塑性的定量link(和其他论文定义一样),真实label和peers的作为监督,构建损失ccl,通过生成构建蒸馏损失
2、SEMANTIC-AWARE REPRESENTATION LEARNING FOR LIFELONG LEARNING
语义相关性,2025icrl,这个思路一样,引导新物体的原型和之前学习到的相似任务,cos相似度
每个物体用物体原型表示,捕捉样本的平均特征,在任务结束后存在memory中。通过评估存的原型的相似度和新物体的平均特征,模型从过去和当前任务识别正负样本。利用这种metric关系引导模型向更相似的原型,
3、Interactive Continual Learning: Fast and Slow Thinking
有训练,也有inference,acquire task related information by leveraging category features and small model knowledge
4、Towards Efficient Replay in Federated Incremental Learning
结合了新任务的,设计全局,表示所有clients的样本
5、Learning to Remember:
ASynaptic Plasticity Driven Framework for Continual Learning
6、Dynamic Generative Memory的可塑性 plasticity(CVPR2019)
Adaptive Visual Scene Understanding: Incremental Scene Graph Generation
保存图中的场景图要素(图中的objects和relations),以此作为prompt生成过去的样本以保存过去的知识
7、Neuromimetic metaplasticity for adaptive continual learning
方法是网络中的突出随机连接,不需要额外训练或结构修改,动态分配旧的和新的资源,在记忆容量和性能之间去平衡,画图可参考
8、Experience Replay Addresses Loss of Plasticity in Continual Learning
2025经验重放、icl,在情景中学习。通过transformer就可以使plas损失消失?好像没看到方法
参考GitHub - xialeiliu/Awesome-Incremental-Learning: Awesome Incremental Learning
更多推荐
所有评论(0)