Class-incremental image classification on TinyImageNet
71.51Last AccuracyGRPO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GRPOTraining Protocol=Joint Training2026.05 | 71.51 | — | — | |
| RaPOTraining Protocol=Continual Learning, Number of Tasks=102026.05 | 62.36 | 16.94 | — | |
| RaPOTraining Protocol=Continual Learning, Number of Tasks=202026.05 | 62.2 | 14.68 | — | |
| GRPOTraining Protocol=Continual Learning, Number of Tasks=102026.05 | 46.84 | 36.24 | — | |
| GRPOTraining Protocol=Continual Learning, Number of Tasks=202026.05 | 42.08 | 40.85 | — | |
| Constrained Expansive Over-Sampling (CEOS)T=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 35.8 | — | 49 | |
| SFTTraining Protocol=Joint Training2026.05 | 34.71 | — | — | |
| EFCT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 34.5 | — | 47.9 | |
| LDCT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 34.2 | — | 46.8 | |
| Constrained Expansive Over-Sampling (CEOS)T=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 31.8 | — | 44.3 | |
| FeTrILT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 31 | — | 45.3 | |
| FeCAMT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 30.9 | — | 44.9 | |
| ADCT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 30.2 | — | 42.4 | |
| PRAKAT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 29.7 | — | 40.7 | |
| SDCT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 29.5 | — | 43.8 | |
| EFCT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 28.4 | — | 42.1 | |
| LwFT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 27.1 | — | 39.6 | |
| SDCT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 26.3 | — | 40.6 | |
| FeTrILT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 25.9 | — | 39.9 | |
| FeCAMT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 24.9 | — | 37.9 | |
| LDCT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 24.9 | — | 38.2 | |
| PASST=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 24.5 | — | 39.5 | |
| PRAKAT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 24.4 | — | 35.6 | |
| Constrained Expansive Over-Sampling (CEOS)T=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 23.2 | — | 36.3 | |
| EFCT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 21.5 | — | 34.8 | |
| EWCT=10, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 18.5 | — | 34.3 | |
| PASST=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 18.5 | — | 30.4 | |
| ADCT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 18.1 | — | 36 | |
| EWCTraining Protocol=Continual Learning, Number of Tasks=102026.05 | 17.17 | 41.55 | — | |
| SFTTraining Protocol=Continual Learning, Number of Tasks=102026.05 | 15.94 | 43.29 | — | |
| LDCT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 15.3 | — | 29.7 | |
| LwFT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 15.2 | — | 31.5 | |
| L2Training Protocol=Continual Learning, Number of Tasks=102026.05 | 15.18 | 44.1 | — | |
| LwFTraining Protocol=Continual Learning, Number of Tasks=102026.05 | 15.07 | 36.41 | — | |
| ADCT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 13.6 | — | 26.8 | |
| LwFTraining Protocol=Continual Learning, Number of Tasks=202026.05 | 12.51 | 45.82 | — | |
| EWCTraining Protocol=Continual Learning, Number of Tasks=202026.05 | 11.41 | 51.02 | — | |
| FeCAMT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 11.4 | — | 24.3 | |
| EWCT=20, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 11.3 | — | 26.8 | |
| SDCT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 10.9 | — | 24.8 | |
| FeTrILT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 10.3 | — | 22.9 | |
| PRAKAT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 9.4 | — | 22.5 | |
| SFTTraining Protocol=Continual Learning, Number of Tasks=202026.05 | 8.71 | 53.68 | — | |
| L2Training Protocol=Continual Learning, Number of Tasks=202026.05 | 8.6 | 53.98 | — | |
| PASST=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=true2026.06 | 7.2 | — | 20.4 | |
| LwFT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 6 | — | 18.3 | |
| EWCT=40, Training from scratch=true, Backbone=ResNet-18, Prototype rehearsal approach=false2026.06 | 3.1 | — | 12.3 | |
| AdaGaussT (Number of tasks)=10, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 18.7 | — | |
| AdaGaussT (Number of tasks)=20, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 23.1 | — | |
| ADCT (Number of tasks)=10, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 30.2 | — | |
| ADCT (Number of tasks)=20, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 36.8 | — | |
| BiCycT (Number of tasks)=10, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 12 | — | |
| BiCycT (Number of tasks)=20, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 18.9 | — | |
| EFCT (Number of tasks)=10, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 23.5 | — | |
| EFCT (Number of tasks)=20, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 30.1 | — | |
| LDCT (Number of tasks)=10, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 24.7 | — | |
| LDCT (Number of tasks)=20, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 30.7 | — | |
| LwFT (Number of tasks)=10, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 21.9 | — | |
| LwFT (Number of tasks)=20, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 33.5 | — | |
| SDCT (Number of tasks)=10, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 25.1 | — | |
| SDCT (Number of tasks)=20, Backbone=ResNet-18, Batch size=256, Training Initialization=from-scratch2026.06 | — | 29.4 | — |