Multi-domain Task-Incremental Learning on MTIL Order I
70.4Transfer AccOurs
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| OursExtra data=×, Params.=30.8 M2026.04 | 70.4 | 79.3 | 88.3 | 79.3 | |
| Ours†Extra data=×, Params.=4.6 M2026.04 | 70 | 78.6 | 87.6 | 78.7 | |
| Zero-shotBackbone=ViT-B/16, Evaluation protocol=Zero-shot CLIP2025.03 | 69.4 | 65.3 | 65.3 | — | |
| Zero-shotExtra data=-, Params.=-2026.04 | 69.4 | 65.3 | 65.3 | 66.7 | |
| GIFTBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K, Source of distillation data=Synthetic (Stable Diffusion v1.5), Volume of distillation data=1K per task2025.03 | 69.3 | 77.3 | 86 | — | |
| GIFTExtra data=✓, Params.=149.6 M2026.04 | 69.3 | 77.3 | 86 | 77.5 | |
| MoE-AdapterBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K2025.03 | 68.9 | 76.7 | 85 | — | |
| MoE-AdapterExtra data=×, Params.=59.6 M2026.04 | 68.9 | 76.7 | 85 | 76.9 | |
| DIKIExtra data=×, Params.=1.8 M2026.04 | 68.7 | 76.3 | 85.1 | 76.7 | |
| ZSCLBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K, Source of distillation data=ImageNet, Volume of distillation data=100K2025.03 | 68.1 | 75.4 | 83.6 | — | |
| ZSCLExtra data=✓, Params.=149.6 M2026.04 | 68.1 | 75.4 | 83.6 | 75.7 | |
| l2 baselineBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K2025.03 | 61 | 62.7 | 75.9 | — | |
| LwF-VRBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K2025.03 | 57.2 | 65.1 | 76.6 | — | |
| LwFBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K2025.03 | 56.9 | 64.7 | 74.6 | — | |
| WiSE-FTBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K2025.03 | 52.3 | 60.7 | 77.7 | — | |
| iCaRLBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K2025.03 | 50.4 | 65.7 | 80.1 | — | |
| Continual FinetuneBackbone=ViT-B/16, Batch size=64, Training iterations per task=1K2025.03 | 44.6 | 55.9 | 77.3 | — |