Task-incremental learning on Split CIFAR-100 (test)
93.89Average Accuracy (A_T)MULTITASK
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MULTITASKLearning Scenario=Standard Continual Learning, Replay Buffer (B)=true2023.06 | 93.89 | — | |
| KD w/ QDILearning Scenario=Standard Continual Learning, Replay Buffer (B)=false2023.06 | 88.3 | 2 | |
| KDLearning Scenario=Standard Continual Learning, Replay Buffer (B)=false2023.06 | 87.13 | 4.26 | |
| KD w/ BUFFERLearning Scenario=Standard Continual Learning, Replay Buffer (B)=true2023.06 | 80.35 | 10.76 | |
| KD w/ QDILearning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=false2023.06 | 76.44 | 12.31 | |
| KD w/ BUFFERLearning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=true2023.06 | 73.98 | 16.4 | |
| DER++ w/ AdaB2NBase Algorithm=DER++, Normalization=AdaB2N, Batch Size |B|=10, Memory Size |M|=50002023.10 | 71.7 | — | |
| DER++ w/ AdaB2NBase Algorithm=DER++, Normalization=AdaB2N, Batch Size |B|=10, Memory Size |M|=20002023.10 | 70.33 | — | |
| DER++Learning Scenario=Standard Continual Learning, Replay Buffer (B)=true2023.06 | 69.03 | 23.06 | |
| KDLearning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=false2023.06 | 68.97 | 21.45 | |
| COLDMemory Size (M)=1k2026.06 | 68.03 | 10 | |
| DERLearning Scenario=Standard Continual Learning, Replay Buffer (B)=true2023.06 | 67.74 | 24.77 | |
| DER++ w/ CNBase Algorithm=DER++, Normalization=CN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 67.39 | — | |
| COLD-ORACLEMemory Size (M)=1k2026.06 | 67.38 | 18 | |
| REFRESH-2024Memory Size (M)=1k2026.06 | 67.34 | 6 | |
| ERLearning Scenario=Standard Continual Learning, Replay Buffer (B)=true2023.06 | 67.06 | 25.31 | |
| ER-ACE w/ AdaB2NBase Algorithm=ER-ACE, Normalization=AdaB2N, Batch Size |B|=10, Memory Size |M|=50002023.10 | 67.01 | — | |
| DER++-2020Memory Size (M)=1k2026.06 | 66.96 | 7 | |
| DER++ w/ BNBase Algorithm=DER++, Normalization=BN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 66.23 | — | |
| DER++ w/ CNBase Algorithm=DER++, Normalization=CN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 66.2 | — | |
| DER++ w/ BNBase Algorithm=DER++, Normalization=BN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 65.53 | — | |
| ER-ACE-2022Memory Size (M)=1k2026.06 | 64.97 | 5 | |
| A-GEMLearning Scenario=Standard Continual Learning, Replay Buffer (B)=true2023.06 | 63.88 | 29.98 | |
| ER-ACE w/ AdaB2NBase Algorithm=ER-ACE, Normalization=AdaB2N, Batch Size |B|=10, Memory Size |M|=20002023.10 | 63.88 | — | |
| ER-ACE w/ BNBase Algorithm=ER-ACE, Normalization=BN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 63.83 | — | |
| CBA-2023Memory Size (M)=1k2026.06 | 63.7 | 3 | |
| A-GEM-2019Memory Size (M)=1k2026.06 | 63.59 | 30 | |
| ER-ACE w/ CNBase Algorithm=ER-ACE, Normalization=CN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 63.42 | — | |
| ER-ACE w/ BNBase Algorithm=ER-ACE, Normalization=BN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 61.21 | — | |
| ER-ACE w/ CNBase Algorithm=ER-ACE, Normalization=CN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 61 | — | |
| DER++Learning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=true2023.06 | 54.81 | 35.2 | |
| DERLearning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=true2023.06 | 53.01 | 37.33 | |
| ERLearning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=true2023.06 | 52.59 | 36.41 | |
| DER-2020Memory Size (M)=1k2026.06 | 50.27 | 22 | |
| ER-ACE w/ INBase Algorithm=ER-ACE, Normalization=IN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 50.15 | — | |
| DILearning Scenario=Standard Continual Learning, Replay Buffer (B)=false2023.06 | 49.64 | 35.29 | |
| ER-ACE w/ INBase Algorithm=ER-ACE, Normalization=IN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 49.46 | — | |
| DER++ w/ INBase Algorithm=DER++, Normalization=IN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 49.17 | — | |
| DER++ w/ INBase Algorithm=DER++, Normalization=IN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 47.39 | — | |
| MER-2019Memory Size (M)=1k2026.06 | 47.35 | 5 | |
| DER++ w/ GNBase Algorithm=DER++, Normalization=GN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 45.12 | — | |
| ER-ACE w/ GNBase Algorithm=ER-ACE, Normalization=GN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 44.84 | — | |
| DER++ w/ LNBase Algorithm=DER++, Normalization=LN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 44.42 | — | |
| ER-ACE w/ LNBase Algorithm=ER-ACE, Normalization=LN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 44.32 | — | |
| DER++ w/ GNBase Algorithm=DER++, Normalization=GN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 43.54 | — | |
| DER++ w/ LNBase Algorithm=DER++, Normalization=LN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 43.24 | — | |
| ER-ACE w/ GNBase Algorithm=ER-ACE, Normalization=GN, Batch Size |B|=10, Memory Size |M|=50002023.10 | 42.5 | — | |
| ER-ACE w/ LNBase Algorithm=ER-ACE, Normalization=LN, Batch Size |B|=10, Memory Size |M|=20002023.10 | 42.29 | — | |
| FINETUNELearning Scenario=Standard Continual Learning, Replay Buffer (B)=false2023.06 | 40.13 | 53.64 | |
| DILearning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=false2023.06 | 38.33 | 47.08 | |
| GDumb-2020Memory Size (M)=1k2026.06 | 37.56 | — | |
| A-GEMLearning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=true2023.06 | 35.37 | 56.41 | |
| EWC-2017Memory Size (M)=1k2026.06 | 35 | 17 | |
| Fine-tuneMemory Size (M)=1k2026.06 | 29.86 | 54 | |
| NCCL-2023Memory Size (M)=1k2026.06 | 29.2 | 54 | |
| FINETUNELearning Scenario=Heterogeneous Continual Learning, Replay Buffer (B)=false2023.06 | 27.08 | 64.54 |