Class-Incremental Learning on VTAB CIL, T=5 (test)
92.2Avg AccuracyRanPAC
Evaluation results
| Method | Links | ||
|---|---|---|---|
| RanPACAlgorithm=12023.07 | 92.2 | 53 | |
| RanPAC (No Phase 1)Phase 1=false2023.07 | 92.2 | — | |
| RanPAC (No RPs or Phase 1)RP layer=false, Phase 1=false2023.07 | 91.9 | — | |
| RanPAC (No RPs)RP layer=false2023.07 | 90 | — | |
| RanPAC (NCM only)Evaluation=NCM, Phase 1=false2023.07 | 88.4 | 4 | |
| RanPAC (NCM with Phase 1)Evaluation=NCM, Phase 1=true2023.07 | 88.2 | 22 | |
| CODA-Prompt2023.07 | 87.4 | — | |
| Joint linear probeJoint training=true2023.07 | 86.7 | — | |
| ADaM2023.07 | 84.3 | — | |
| DualPrompt2023.07 | 81.2 | — | |
| L2P2023.07 | 77.1 | — |