Table 2 Correlation between test area-under-the-curve and sample size above and below auto-encoder loss inflection points for with varying dataset characteristics and neural network hyper-parameters.

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

  

Pre-MCSE

Post-MCSE

Parameter

Value

R2

Kendall’s τ

Spearman’s ρ

R2

Kendall’s τ

Spearman’s ρ

N-informative

8

0.074

0.292

0.341

0.577

0.837

0.944

 

16

0.019

−0.172

−0.210

0.578

0.836

0.944

 

32

0.013

−0.180

−0.224

0.534

0.850

0.954

 

64

0.078

0.172

0.212

0.529

0.835

0.944

 

128

0.056

0.138

0.170

0.632

0.804

0.925

 

256

0.010

0.081

0.105

0.729

0.707

0.852

N-classes

2

0.000

0.000

0.000

0.161

0.399

0.530

 

4

0.011

0.072

0.093

0.332

0.514

0.679

 

6

0.039

0.146

0.189

0.522

0.610

0.783

 

8

0.029

0.125

0.162

0.615

0.659

0.824

 

10

0.023

0.118

0.152

0.606

0.617

0.781

N-features

256

0.004

0.059

0.053

0.681

0.585

0.736

 

512

0.149

0.410

0.538

0.666

0.663

0.814

 

784

0.020

−0.108

−0.141

0.713

0.697

0.849

 

1024

0.225

0.402

0.475

0.566

0.624

0.789

 

2048

0.100

0.193

0.253

0.629

0.608

0.769

 

4096

0.069

0.181

0.225

0.582

0.513

0.671

Hidden Layer Size

64

0.012

0.085

0.110

0.459

0.499

0.653

 

128

0.023

0.118

0.153

0.552

0.591

0.755

 

256

0.014

0.085

0.111

0.590

0.619

0.783

 

512

0.023

0.109

0.142

0.621

0.638

0.800

 

784

0.026

0.123

0.160

0.603

0.656

0.818

 

1024

0.039

0.151

0.194

0.595

0.640

0.805

Back to article page