2024年秋季学期
2024.12.17
- 郑金鹏：CLIPCleaner Cleaning Noisy Labels with CLIP[paper][slides]
2024.12.10
- 郑腾鑫陵：Efficient Test-Time Adaptation of Vision-Language Models & WATT: Weight Average Test-Time Adaptation of CLIP[paper1[paper2][slide]
第一篇文章使用CLIP为backbone，不调节CLIP参数，设计positive和negative队列作为缓存模型，减少时间开销。第二篇使用CLIP为backbone，根据text template 调整CLIP参数，验证不同text template的影响。
- 卢昕怡：Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning [paper][slides]
2024.12.3
- 赵世佶：Decorate the Newcomers: Visual Domain Prompt for Continual Test Time Adaptation & Dataset Condensation with Gradient Matching [paper1] [paper2] [slide]
第一篇文章做的是持续TTA的设定。文章通过直接在图像像素层面构造矩形prompt作为可变参数，并和输入图像聚合送入冻结的模型中得到输出。作者设计了两种prompt，一个用教师-学生网络，通过交叉熵更新学习域知识；另一个多了一项正则限制域敏感参数的更新以保证域泛化知识保留。第二篇是数据集蒸馏的经典论文，通过课程梯度匹配的方式实现数据集蒸馏。
2024.11.26
-
陶子健： Towards Calibrated Multi-label Deep Neural Networks[paper][slides]
-
郑金鹏：GLAD: Towards Better Reconstruction with Global and Local Adaptive Diffusion Models for Unsupervised Anomaly Detection[paper][slides]
2024.11.19
- 卢昕怡：Towards Unbiased Training in Federated Open-world Semi-supervised Learning [paper] [slide]
- 郑腾鑫陵：A Versatile Framework for Continual Test-Time Domain Adaptation: Balancing Discriminability and Generalizability
[paper][silde]
2024.11.12
- 赵世佶：Self-supervised Learning for Large-scale Item Recommendations & Deep Interest Network for Click-Through Rate Prediction [paper1] [paper2] [slide]
简要介绍了推荐系统的基础架构：召回->粗排->精排->重排，然后分别展示了召回和精排的两个经典模型。谷歌在召回双塔模型的基础上引入对比学习扩充了样本量，解决了长尾表征的问题。阿里利用attention对用户序列建模，设计了DIN。
- 郑宇祥：Learning Equi-angular Representations for Online Continual Learning[paper][slides]
2024.11.6
- 郑金鹏：AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models[paper][slides]
2024.10.29
- 卢昕怡：Robust Semi-supervised Learning by wisely leveraging open-set data [paper] [slide]
- 郑宇祥：Breaking the Memory Barrier: Near Infinite Batch Size Scaling for Contrastive Loss[paper][slides]
2024.10.21
- 赵世佶：EcoTTA: Memory-Efficient Continual Test-time Adaptation via Self distilled Regularization [paper] [slide]
介绍了EcoTTA，一篇研究持续TTA过程中如何减少内存消耗，从而能在端侧设备上实现持续适应。文章借鉴了TinyTL、EATA等一系列工作，通过将源模型的encoder分块，并在每一块之间插入专属的元网络(包括一层BN和一个Conv的残差连接)，实现模型激活值内存占用大幅缩减。设计的损失函数包括经典的熵最小化，以及一个子蒸馏的正则项用于防止模型遗忘。
- 陶子健：Pi-DUAL: Using privileged information to distinguish clean from noisy labels [paper] [slide]
2024.10.14
- 郑腾鑫陵：Continual test-time domain Adaption [paper] [slide]
2024.9.23
- 郑金鹏：DeiT-LT: Distillation Strikes Back for Vision Transformer Training on Long-Tailed Datasets [paper] [slide]
- 陶子健：Category-Prompt Refined Feature Learning for Long-Tailed Multi-Label Image Classification [paper] [slide]
2024.9.9
- 卢昕怡： FedCorr Multi-Stage Federated Learning for Label Noise Correction [paper] [slide]
- 郑腾鑫陵：Continual-MAE Adaptive Distribution Masked Autoencoders for Continual Test-Time Adaptation [paper] [slide]
2024年春季学期
写作参考材料