notesum.ai
Published at May 7Visual Prompt Tuning in Null Space for Continual Learning
NeurIPS
Released Date: May 7, 2024
Authors: Yue Lu1, Shizhou Zhang1, De Cheng2, Yinghui Xing1, Nannan Wang2, PENG WANG1, Yanning Zhang1
Aff.: 1School of Computer Science, Northwestern Polytechnical University, China; 2School of Telecommunications Engineering, Xidian University, China
Arxiv: https://openreview.net/pdf/bdbd2120af155b8c2cb5b5408d55ce8714e2d9eb.pdf

| Method | 10S-CIFAR-100 | 20S-CIFAR-100 | 10S-ImageNet-R | 10S-DomainNet | ||||
|---|---|---|---|---|---|---|---|---|
| Acc. | Forgetting | Acc. | Forgetting | Acc. | Forgetting | Acc. | Forgetting | |
| VPT-Seq | 87.27 | 12.33 | 82.36 | 17.36 | 72.46 | 19.41 | 73.28 | 25.65 |
| VPT-NSP2 | 91.74 | 3.28 | 89.89 | 4.91 | 78.88 | 5.06 | 83.54 | 8.54 |
| Upper-bound | 93.87 | - | 93.87 | - | 84.60 | - | 89.25 | - |
| CLIP-Seq | 72.91 | 15.13 | 71.37 | 17.89 | 75.69 | 19.21 | 67.73 | 35.60 |
| CLIP-NSP2 | 80.96 | 12.45 | 79.83 | 13.77 | 82.17 | 6.42 | 77.04 | 18.33 |
| Upper-bound | 84.52 | - | 84.52 | - | 84.86 | - | 81.65 | - |