Research map / Continual learning / Class-incremental
CroMo-Mixup: Augmenting Cross-Model Representations for Continual Self-Supervised Learning
Erum Mushtaq, Duygu Nur Yaldiz, Yavuz Faruk Bakman, Jie Ding, Chenyang Tao, Dimitrios Dimitriadis, Salman Avestimehr
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-task data mixup and cross-model feature mixup to CSSL | CaSSLe+ (Barlow-Twins) · CIFAR100-Split5 | average linear accuracy: 61.3 -> 65.5 | Improved |
| add CroMo-Mixup on CIFAR100-Split10 | Barlow-Twins CSSL · CIFAR100-Split10 | linear accuracy: 56.64 -> 62.48 | Improved |
| add CroMo-Mixup on CIFAR10-Split2 with Barlow-Twins | Barlow-Twins CSSL · CIFAR10-Split2 | Task-ID prediction: 95.87 -> 95.75 | No clear effect |
| split classes across tasks in minibatch training instead of uniform sampling (ablation) | contrastive SSL baselines · CIFAR100 task confusion probe | linear accuracy and task-ID prediction: drop of 4% or more | Hurt |
| split classes across tasks in supervised training (ablation) | supervised learning · CIFAR100 task confusion probe | linear accuracy | No clear effect |
| split data (not classes) across tasks in SSL training (ablation) | contrastive SSL baselines · CIFAR100 data-incremental probe | linear accuracy | No clear effect |
| use cross-task instead of within-task data mixup (ablation) | CroMo-Mixup · CIFAR100-Split5 (Barlow-Twins) | accuracy: 54.16 -> 62.31 | Improved |
| use cross-model instead of same-model feature mixup (ablation) | CroMo-Mixup · CIFAR100-Split5 (Barlow-Twins) | accuracy: 62.31 -> 63.94 | Improved |
| add distillation with temporal projector (ablation) | CroMo-Mixup · CIFAR100-Split5 (Barlow-Twins) | accuracy: 63.94 -> 65.48 | Improved |
| reduce memory buffer to 25 samples per task (ablation) | CroMo-Mixup · CIFAR100-Split10 | average linear accuracy: drops of 1-2 percentage points | Hurt |
| apply CroMo-Mixup to out-of-distribution transfer | ResNet-18 trained on CIFAR100-Split5 · Flower102 | linear accuracy: 44.17 -> 49.03 | Improved |
Closest papers by meaning
- Instance and Category Supervision are Alternate Learners for Continual Learning (ICCV 2023 · similarity 0.89)
- Representational Continuity for Unsupervised Continual Learning (ICLR 2022 · similarity 0.88)
- Probing Representation Forgetting in Supervised and Unsupervised Continual Learning (CVPR 2022 · similarity 0.88)
- Continual Learning with Global Alignment (NeurIPS 2024 · similarity 0.87)
- AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection (NeurIPS 2025 · similarity 0.86)
- Continual Learning Using a Kernel-Based Method Over Foundation Models (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
