A comparison of ANN and SVM model performance on the two-spiral and modified two-spiral classification tasks.

A feed-forward ANN using PyTorch with layer dimensions: 2-16-32-1; Learning rate: 0.1

Network Architecture:

Sequential (
(0): Linear (in_features=2, out_features=16, bias=True)
(1): ELU (alpha=1.0)
(2): Linear (in_features=16, out_features=32, bias=True)
(3): ELU (alpha=1.0)
(4): Linear (in_features=32, out_features=1, bias=True)
(5): Sigmoid () )

ANN decision boundary after 600 iterations:

SVM decision boundary:

A feed-forward ANN using PyTorch with layer dimensions: 2-16-32-1; Learning rate: 0.1

Network Architecture:

Sequential (
(0): Linear (in_features=2, out_features=16, bias=True)
(1): ELU (alpha=1.0)
(2): Linear (in_features=16, out_features=32, bias=True)
(3): ELU (alpha=1.0)
(4): Linear (in_features=32, out_features=1, bias=True)
(5): Sigmoid () )

Comparison SVM parameters with decision boundaries visualised: