Table 5 Temperature-schedule ablation on ImageNet (ResNet-50), the bold values are the best results

From: A dynamic fractional generalized deterministic annealing for rapid convergence in deep learning optimization

Schedule

Epochs to 80 %

Final top-1 (%)

Relative speed

Entropy-controlled (ours)

62

79.4

-

Geometric (γ = 0.95)

99

78.7

1.6 × slower

Fixed (\(T={T}_{\max }\))

147

78.2

2.4 × slower

Back to article page