Table 5 Accuracy per layer and statistical features of their filters for EfficientNet-B0 trained on \(K\) labels from CIFAR-100.

From: Towards a universal mechanism for successful deep learning

Stage

\({N}_{f}\)

\({F}_{s}\)

\(F{C}_{s}\)

Accuracy

\(n\)

\({N}_{c}\)

\({C}_{s}\)

EfficientNet-B0 on CIFAR-10/100

 9

1280

1 × 1

1280

0.986

3.8

1.08

1.6

 7

192

1 × 1

192

0.955

8.3

1.80

1.3

 5

80

2 × 2

320

0.851

10.6

1.85

1.2

 4

40

4 × 4

640

0.845

12.8

2.15

1.3

 3

24

8 × 8

1536

0.755

14.5

2.75

1.3

 1

32

16 × 16

8192

0.634

18.1

1.55

1.9

EfficientNet-B0 on CIFAR-20/100

 9

1280

1 × 1

1280

0.973

8.1

1.1

2.0

 7

192

1 × 1

192

0.915

22.9

2.1

1.5

 5

80

2 × 2

320

0.765

29.7

2.0

1.3

 4

40

4 × 4

640

0.764

40.8

2.8

1.4

 3

24

8 × 8

1536

0.645

48.1

3.4

1.4

 1

32

16 × 16

8192

0.482

63.8

1.5

3.1

EfficientNet-B0 on CIFAR-40/100

 9

1280

1 × 1

1280

0.935

16.4

1.1

2.4

 7

192

1 × 1

192

0.849

64.2

2.6

1.7

 5

80

2 × 2

320

0.652

85.2

2.7

1.4

 4

40

4 × 4

640

0.650

111.3

3.3

1.5

 3

24

8 × 8

1536

0.553

129.6

3.9

1.6

 1

32

16   × 16

8192

0.362

223.6

1.9

3.9

EfficientNet-B0 on CIFAR-60/100

 9

1280

1 × 1

1280

0.915

21.4

1.2

2.6

 7

192

1 × 1

192

0.810

121.8

3.2

1.9

 5

80

2 × 2

320

0.593

152.0

3.0

1.6

 4

40

4 × 4

640

0.603

200.0

3.8

1.6

 3

24

8 × 8

1536

0.511

252.3

4.8

1.7

 1

32

16 × 16

8192

0.313

492.6

2.5

4.4

  1. The results here are similar to those of Table 2, where EfficientNet-B0 was trained on \(K=10, 20, 30,\) and \(60\) labels out of 100, namely CIFAR-K/100 (Supplementary Information).
Back to article page