Continual Learning on Permuted-MNIST (Mean test accuracy)
98.6Mean Test AccuracyHAT
Evaluation results
| Method | Links | |
|---|---|---|
| HATReplay=false, Single Epoch=false2021.06 | 98.6 | |
| Multi-Head (multi-task)2021.06 | 98.16 | |
| IsolatedReplay=false, Single Epoch=false2021.06 | 98.03 | |
| RMNReplay=false, Single Epoch=false2021.06 | 97.73 | |
| Model ZooReplay=true, Single Epoch=false2021.06 | 97.71 | |
| SIReplay=false, Single Epoch=false2021.06 | 97.1 | |
| EWCReplay=false, Single Epoch=false2021.06 | 96.9 | |
| VCLReplay=true, Single Epoch=false2021.06 | 95.5 | |
| FROMPReplay=true, Single Epoch=false2021.06 | 94.9 | |
| FRCLReplay=true, Single Epoch=false2021.06 | 94.3 | |
| Prog-NetsReplay=false, Single Epoch=false2021.06 | 93.5 | |
| MEGA-IIReplay=true2021.06 | 91.2 | |
| A-GEMReplay=true, Single Epoch=true2021.06 | 89.1 | |
| OGDReplay=false, Single Epoch=true2021.06 | 86.44 | |
| MC-SGDReplay=true, Single Epoch=true2021.06 | 85.3 | |
| Layerwise Optimization by Gradient DecompositionTraining Epochs=3 epochs2021.05 | 84.3 | |
| GEMTraining Epochs=3 epochs, Memory Size=2562021.05 | 84.1 | |
| A-GEMTraining Epochs=3 epochs, Memory Size=2562021.05 | 82.9 | |
| Layerwise Optimization by Gradient DecompositionTraining Epochs=1 epoch2021.05 | 82.9 | |
| GEMReplay=true, Single Epoch=true2021.06 | 82.6 | |
| tinyER+Training Epochs=1 epoch2021.05 | 82.5 | |
| S-GEMTraining Epochs=3 epochs, Memory Size=2562021.05 | 82.3 | |
| GEMTraining Epochs=1 epoch, Memory Size=2562021.05 | 81.9 | |
| Stable-SGDReplay=false, Single Epoch=true2021.06 | 80.1 | |
| ER-ReservoirReplay=true, Single Epoch=true2021.06 | 79.8 | |
| A-GEMTraining Epochs=1 epoch, Memory Size=2562021.05 | 76.4 | |
| S-GEMTraining Epochs=1 epoch, Memory Size=2562021.05 | 76.3 | |
| GDumbTraining Epochs=1 epoch, Training Schedule=Online Task Incremental Disjoint2021.05 | 73.4 | |
| MultimodelTraining Epochs=1 epoch2021.05 | 71.9 | |
| EWCTraining Epochs=3 epochs2021.05 | 61.9 | |
| IndependentTraining Epochs=3 epochs2021.05 | 61.3 | |
| MultimodelTraining Epochs=3 epochs2021.05 | 59.4 | |
| EWCTraining Epochs=1 epoch2021.05 | 57.5 | |
| SingleTraining Epochs=1 epoch2021.05 | 57.4 | |
| SingleTraining Epochs=3 epochs2021.05 | 53 | |
| IndependentTraining Epochs=1 epoch2021.05 | 37 | |
| RMN2021.02 | 0.9773 | |
| FROMPData replay buffer=true, Points per task=402021.02 | 0.949 | |
| FRCLData replay buffer=true, Points per task=2002021.02 | 0.943 | |
| HATPretrained network=true, Multiheaded=true2021.02 | 0.916 | |
| MEGA-IIData replay buffer=true, Trained over all tasks=false, Points per task=2562021.02 | 0.9121 | |
| VCLData replay buffer=true, Multiheaded=true, Points per task=2002021.02 | 0.9 | |
| EWC2021.02 | 0.84 | |
| DLP2021.02 | 0.82 |