Research map / Vision-language and multimodal / CLIP and zero-shot
Pi-CCA: Prompt-Invariant CCA Certificates for Replay-Free Continual Multimodal Learning
Jiayu Zhang, Chuangxin Zhao, Canran Xiao, Ruibo Duan, Wenyi Mo, Haoyu Gao, Wenshuo Wang
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 spectral preservation term (ablation) | PI-CCA · MTIL | Avg accuracy: 74.3 -> 76.8 | Improved |
| add subspace-angle term (ablation) | PI-CCA · MTIL | Avg accuracy: 74.6 -> 76.8 | Improved |
| add prompt invariance loss (ablation) | PI-CCA · MTIL | Avg accuracy: 75.3 -> 76.8 | Improved |
| add certificate EMA (ablation) | PI-CCA · MTIL | Avg accuracy: 75.6 -> 76.8 | Improved |
| add streaming covariance EMA (ablation) | PI-CCA · MTIL | Avg accuracy: 74.1 -> 76.8 | Improved |
| add low-order spectral moments (J>0) (ablation) | PI-CCA · MTIL | Avg accuracy: 76.1 -> 76.8 | Improved |
| use sorted surrogate instead of exact Hungarian pairing (ablation) | PI-CCA · MTIL | Avg accuracy: 76.7 -> 76.8 | No clear effect |
| Gaussian instead of SRHT sketches (ablation) | PI-CCA · MTIL | Avg accuracy: 76.6 -> 76.8 | No clear effect |
| stop-gradient certificate update instead of backprop through EMA+QR (ablation) | PI-CCA · MTIL | Avg accuracy: 76.0 -> 76.8 | Improved |
| full anchor set instead of 80% reduced anchors (ablation) | PI-CCA · MTIL | Avg accuracy: 75.4 -> 76.8 | Improved |
| sketched projector loss instead of explicit principal-angle loss (ablation) | PI-CCA · MTIL | Avg accuracy: 76.9 -> 76.8 | No clear effect |
| larger backbone ViT-L/14 instead of ViT-B/16 | PI-CCA · MTIL | Avg accuracy: 76.8 -> 78.2 | Improved |
Closest papers by meaning
- Subspace Alignment for CLIP-based Continual Learning via Canonical Correlation Analysis (CVPR 2026 · similarity 0.88)
- Reversible Primitive–Composition Alignment for Continual Vision–Language Learning (ICLR 2026 · similarity 0.87)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.86)
- Reconstruction Alignment Improves Unified Multimodal Models (ICLR 2026 · similarity 0.85)
- Advancing Prompt-Based Methods for Replay-Independent General Continual Learning (ICLR 2025 · similarity 0.85)
- Context-Aware Multimodal Pretraining (CVPR 2025 · similarity 0.85)
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.
