List your company
ResearchFormerly WizWand

Continual Learning on {MNIST, SVHN, notMNIST, FashionMNIST, CIFAR-10} (test)

90.56AADECODE
10.053630.954351.85572.7557Nov 7, 2024Nov 10, 2024Nov 13, 2024Nov 16, 2024Nov 19, 2024Nov 22, 2024Nov 26, 2024
Updated 5mo ago

Evaluation results

MethodLinks
DECODEBackbone=ViT, Optimization Epochs=252024.1190.56—0.05
DualpromptBackbone=ViT, Optimization Epochs=252024.1188.08—2.21
PEFT Ensemble (FiLM)Differential Privacy budget (epsilon)=8, Privacy parameter (delta)=1e-5, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1187.833—
DER++Memory Buffer Size=500, Backbone=ViT, Optimization Epochs=252024.1184.88—10.46
Coda-PBackbone=ViT, Optimization Epochs=252024.1184.82—13.24
ERMemory Buffer Size=500, Backbone=ViT, Optimization Epochs=252024.1184.26—12.85
F2PBackbone=ViT, Optimization Epochs=252024.1181.14—4.64
PEFT Ensemble (FiLM)Differential Privacy budget (epsilon)=1, Privacy parameter (delta)=1e-5, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1179.695—
PEFT Ensemble (FiLM)Differential Privacy budget (epsilon)=non-DP, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1165.7514—
Cosine classifierDifferential Privacy budget (epsilon)=non-DP, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1159.870—
Cosine classifierDifferential Privacy budget (epsilon)=8, Privacy parameter (delta)=1e-5, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1159.780—
Cosine classifierDifferential Privacy budget (epsilon)=1, Privacy parameter (delta)=1e-5, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1158.541—
EWCBackbone=ViT, Optimization Epochs=252024.1150.93—34.94
LwFBackbone=ViT, Optimization Epochs=252024.1147.91—38.01
NaiveDifferential Privacy budget (epsilon)=8, Privacy parameter (delta)=1e-5, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1117.5690—
NaiveDifferential Privacy budget (epsilon)=1, Privacy parameter (delta)=1e-5, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1115.9385—
NaiveDifferential Privacy budget (epsilon)=non-DP, Backbone=ViT-B, Pre-training dataset=ImageNet-21k2024.1113.1579—