Research map / Continual learning / Continual fine-tuning of LLMs
COBRA: A Continual Learning Approach to Vision-Brain Understanding
Xuan-Bac Nguyen, Manuel Serna-Aguilera, Arabinda K. Choudhary, Pawan Sinha, Xin Li, Ky Luu
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 COBRA continual learning (SC + PSS + MRIFormer) versus no continual learning | COBRA · NSD rehearsal-free reconstruction (4 steps, all) | SSIM / CLIP: 0.207 -> 0.328 SSIM, 59.8 -> 93.2 CLIP | Improved |
| use COBRA instead of PLOP in rehearsal-free setting | continual learning method · NSD rehearsal-free reconstruction (4 steps, all) | SSIM / CLIP: 0.288 -> 0.328 SSIM, 76.0 -> 93.2 CLIP | Improved |
| combine COBRA with PLOP | COBRA · NSD rehearsal-free reconstruction (4 steps, all) | SSIM / CLIP: 0.328 -> 0.338 SSIM, 93.2 -> 94.2 CLIP | Improved |
| combine COBRA with PLOP | COBRA · NSD rehearsal-free reconstruction (2 steps, all) | SSIM / CLIP: 0.329 -> 0.333 SSIM, 92.9 -> 93.7 CLIP | Improved |
| use COBRA instead of LwF in 2-step rehearsal-free setting | continual learning method · NSD rehearsal-free reconstruction (2 steps, all) | SSIM / CLIP: 0.249 -> 0.329 SSIM, 64.3 -> 92.9 CLIP | Improved |
| increase rehearsal buffer size from 500 to 4000 samples per subject (ablation) | COBRA · NSD rehearsal-based reconstruction (4 steps, all) | SSIM / CLIP: 0.334 -> 0.343 SSIM, 93.6 -> 95.8 CLIP | Improved |
| use COBRA instead of Co2L in rehearsal-based setting | continual learning method · NSD rehearsal-based reconstruction (4 steps, buffer 500, all) | SSIM / CLIP: 0.300 -> 0.334 SSIM, 83.1 -> 93.6 CLIP | Improved |
| use full-brain fMRI signals instead of partial ROI signals as input (ablation) | COBRA · NSD reconstruction | SSIM / CLIP: 3% higher SSIM and 2% higher CLIP | Improved |
| use MRIFormer transformer encoder-decoder instead of linear layers (ablation) | COBRA · NSD reconstruction | SSIM / CLIP: approximately 3% in both SSIM and CLIP | Improved |
| set PSS prompt top-k to 30 instead of larger or smaller values (ablation) | PSS module · NSD reconstruction | SSIM / CLIP | Improved |
| select subject-specific tokens by prompt-based PSS instead of attention pooling (ablation) | PSS module · Subject identification accuracy | accuracy: 82.8 -> 98.5 | Improved |
| select tokens by prompt-based PSS instead of random (ablation) | PSS module · Subject identification accuracy | accuracy: 21.6 -> 98.5 | Improved |
Closest papers by meaning
- Bisecle: Binding and Separation in Continual Learning for Video Language Understanding (NeurIPS 2025 · similarity 0.86)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.85)
- Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster (AAAI 2024 · similarity 0.85)
- Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning (NeurIPS 2026 · similarity 0.85)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.84)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.84)
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.
