Radial Basis Function Neural Network

Class A — inner Class B — outer

Classifier mode

Classifier accuracy

—

RBF parameters

Centers (K) 9

Width (σ) 0.44

Show Gaussian footprints

Latent space view

In RBFNN mode, the K pre-fixed centers are projected to 2D via PCA of the K-dimensional activation space, and the boundary uses a least-squares solver. In SVM mode, all N=90 training points are considered as candidate centers; the max-margin solver retains only the S support vectors (K snaps to S in the slider), discarding the rest. The feature-space canvas and accuracy bar always reflect the active mode.

Single Layer Perceptron (SLP)

No hidden layer — direct weighted sum

f(z) = sign(z) binary classification (hard threshold; non-differentiable)
f(z) = 1/(1+e⁻ᶻ) Bernoulli / binary probability (differentiable — use for training)
f(z) = exp(z) Poisson / Gamma regression (log link)
f(z) = z linear (Gaussian) regression

Radial Basis Function Network (RBFNN)

Gaussian hidden layer lifts data to separable space

f(z) = sign(z) binary classification (hard threshold; non-differentiable)
f(z) = 1/(1+e⁻ᶻ) Bernoulli / binary probability (differentiable — use for training)
f(z) = exp(z) Poisson / Gamma regression (log link)
f(z) = z linear (Gaussian) regression