Continual Learning on Split-CIFAR-100 5 tasks x 20 classes (train)
43.1Average Accuracy (AA)Round-Robin
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Round-RobinBackbone=ImageNet-pretrained ResNet-18 (frozen), LoRA rank=8, Capacity k/N=0.5, Compression ρ=0.52026.06 | 43.1 | 0.045 | |
| NSR-RL (variant)Backbone=ImageNet-pretrained ResNet-18 (frozen), LoRA rank=8, Capacity k/N=0.5, Compression ρ=0.52026.06 | 40.5 | 0.068 | |
| EWC-InspiredBackbone=ImageNet-pretrained ResNet-18 (frozen), LoRA rank=8, Capacity k/N=0.5, Compression ρ=0.52026.06 | 40.4 | 0.157 | |
| GradientBackbone=ImageNet-pretrained ResNet-18 (frozen), LoRA rank=8, Capacity k/N=0.5, Compression ρ=0.52026.06 | 40.2 | 0.151 | |
| RandomBackbone=ImageNet-pretrained ResNet-18 (frozen), LoRA rank=8, Capacity k/N=0.5, Compression ρ=0.52026.06 | 38.8 | 0.085 | |
| PLANBackbone=ImageNet-pretrained ResNet-18 (frozen), LoRA rank=8, Capacity k/N=0.5, Compression ρ=0.52026.06 | 38.4 | 0.168 | |
| Full-ActiveBackbone=ImageNet-pretrained ResNet-18 (frozen), LoRA rank=8, Capacity k/N=0.5, Compression ρ=0.52026.06 | 37.3 | 0.198 | |
| PEARLBackbone=ImageNet-pretrained ResNet-18 (frozen), LoRA rank=8, Capacity k/N=0.5, Compression ρ=0.52026.06 | 26.2 | 0.247 |