List your company
ResearchFormerly WizWand

Image Classification on CIFAR10 (test) (Error Rate)

0.116Error RateEM routing
-0.566564.040728.64813.25528Feb 11, 2018Jun 20, 2018Oct 28, 2018Mar 7, 2019Jul 15, 2019Nov 22, 2019Mar 31, 2020
Updated 8mo ago

Evaluation results

MethodLinks
EM routingResolution=32x32, Architecture=4 capsule layers, Ensemble size=52018.050.116
FREMResolution=32x32, Architecture=4 capsule layers, Ensemble size=52018.050.143
FRMSResolution=32x32, Architecture=4 capsule layers, Ensemble size=52018.050.156
Baseline CNNResolution=32x32, Architecture=4 capsule layers, Ensemble size=52018.050.192
UniformAugment (UA)Model=SS (26 2x96d)2020.031.9
AutoAugment (AA)Model=SS (26 2x96d)2020.032
PBAModel=SS (26 2x96d)2020.032
Fast AutoAugment (FAA)Model=SS (26 2x96d)2020.032
RandAugment (RA)Model=SS (26 2x96d)2020.032
ProxylessNAS + cutoutSearch Cost (GPU-days)=4, Search Method=gradient-based2019.072.08
UniformAugment (UA)Model=SS (26 2x32d)2020.032.49
P-DARTS + cutoutParams (M)=3.4, Search Cost (GPU-days)=0.3, Search Method=gradient-based2019.072.5
AutoAugment (AA)Model=SS (26 2x32d)2020.032.5
PBAModel=SS (26 2x32d)2020.032.5
Fast AutoAugment (FAA)Model=SS (26 2x32d)2020.032.5
AmoebaNet-B + cutoutParams (M)=2.8, Search Cost (GPU-days)=3150, Search Method=evolution2019.072.55
PC-DARTS + cutoutParams (M)=3.6, Search Cost (GPU-days)=0.1, Search Method=gradient-based2019.072.57
AutoAugment (AA)Model=WRN-28-102020.032.6
PBAModel=WRN-28-102020.032.6
CutoutModel=SS (26 2x96d)2020.032.6
NASNet-A + cutoutParams (M)=3.3, Search Cost (GPU-days)=1800, Search Method=RL2019.072.65
UniformAugment (UA)Model=WRN-28-102020.032.67
Fast AutoAugment (FAA)Model=WRN-28-102020.032.7
RandAugment (RA)Model=WRN-28-102020.032.7
DARTS (2nd order) + cutoutParams (M)=3.3, Search Cost (GPU-days)=1, Search Method=gradient-based2019.072.76
BayesNAS + cutoutParams (M)=3.4, Search Cost (GPU-days)=0.2, Search Method=gradient-based2019.072.81
SNAS (moderate) + cutoutParams (M)=2.8, Search Cost (GPU-days)=1.5, Search Method=gradient-based2019.072.85
ENAS + cutoutParams (M)=4.6, Search Cost (GPU-days)=0.5, Search Method=RL2019.072.89
BaselineModel=SS (26 2x96d)2020.032.9
DARTS (1st order) + cutoutParams (M)=3.3, Search Cost (GPU-days)=0.4, Search Method=gradient-based2019.073
CutoutModel=SS (26 2x32d)2020.033
CutoutModel=WRN-28-102020.033.1
PNASParams (M)=3.2, Search Cost (GPU-days)=225, Search Method=SMBO2019.073.41
DenseNet-BCParams (M)=25.6, Search Method=manual2019.073.46
Drop-ActivationBackbone=DenseNet190-402018.113.51
NAONet-WSParams (M)=3.1, Search Cost (GPU-days)=0.4, Search Method=NAO2019.073.53
Fast AutoAugment (FAA)Model=WRN-40-22020.033.6
BaselineModel=SS (26 2x32d)2020.033.6
AutoAugment (AA)Model=WRN-40-22020.033.7
Drop-ActivationBackbone=WideResNet28-102018.113.74
Hireachical EvolutionParams (M)=15.7, Search Cost (GPU-days)=300, Search Method=evolution2019.073.75
UniformAugment (UA)Model=WRN-40-22020.033.75
RReLUBackbone=DenseNet190-402018.113.84
BaselineBackbone=WideResNet28-102018.113.85
BaselineModel=WRN-28-102020.033.9
BaselineBackbone=DenseNet190-402018.113.91
BaselineBackbone=ResNeXt29-8x642018.113.95
Drop-ActivationBackbone=ResNeXt29-8x642018.113.95
DenseNetDepth=100, # params=7.0M2019.054.1
CutoutModel=WRN-40-22020.034.1
OSNetDepth=210, # params=4.6M2019.054.18
RReLUBackbone=WideResNet28-102018.114.32
Drop-ActivationBackbone=DenseNet100-122018.114.38
OSNetDepth=78, # params=2.2M2019.054.41
RReLUBackbone=ResNeXt29-8x642018.114.56
BaselineBackbone=DenseNet100-122018.114.73
Wide ResNetDepth=16, # params=11.0M2019.054.81
Drop-ActivationBackbone=PreResNet1642018.114.87
pre-act ResNetDepth=1001, # params=10.2M2019.054.92
Wide ResNetDepth=40, # params=8.9M2019.054.97
BaselineBackbone=PreResNet1642018.115.01
RReLUBackbone=DenseNet100-122018.115.06
RReLUBackbone=PreResNet1642018.115.17
DenseNetDepth=40, # params=1.0M2019.055.24
pre-act ResNetDepth=164, # params=1.7M2019.055.46
BaselineModel=WRN-40-22020.035.6
Drop-ActivationBackbone=ResNet1642018.115.62
BaselineBackbone=ResNet1642018.115.94
RReLUBackbone=ResNet1642018.116.08
Drop-ActivationBackbone=ResNet1102018.116.25
Drop-ActivationBackbone=VGG19(BN)2018.116.38
BaselineBackbone=VGG19(BN)2018.116.56
RReLUBackbone=VGG19(BN)2018.116.6
BaselineBackbone=ResNet1102018.116.77
RReLUBackbone=ResNet1102018.117.37
NASBOTSearch budget=10 hours, Training iterations=60K2018.0212.09
RANDSearch budget=10 hours, Training iterations=60K2018.0213.42
EASearch budget=10 hours, Training iterations=60K2018.0214.11
TreeBOSearch budget=10 hours, Training iterations=60K2018.0215.33
VGG-19Training iterations=60K2018.0217.18