Table 1 Comparison of statistical power above and below auto-encoder loss inflection points for various image datasets.

From: Autoencoders for sample size estimation for fully connected neural network classifiers

 

Pre-MCSE

Post-MCSE

Dataset

R2

Kendall’s τ

Spearman’s ρ

R2

Kendall’s τ

Spearman’s ρ

MNIST

0.449

0.438

0.611

0.751

0.851

0.972

FMNIST

0.203

0.286

0.411

0.922

0.855

0.969

KMNIST

0.129

0.284

0.414

0.919

0.839

0.971

EMNIST

0.119

0.224

0.336

0.784

0.843

0.971

QMNIST

0.207

0.323

0.476

0.779

0.836

0.962

CIFAR10

0.028

0.051

0.073

0.499

0.777

0.936

STL10

0.002

0.013

0.016

0.002

0.144

0.209

FAKE

0.002

0.085

0.128

0.154

0.032

0.046

Back to article page