Extreme Learning Machine

Wikipedia Description

To predict $\mathbf y \vert \mathbf x$:

1. Make random matrix $\mathbf W_1$
2. Return the least squares estimate with feature vector $\sigma\left(\mathbf W_1 \mathbf x\right)$, where $\sigma$ is an activation function.

Turns out this has already been done here. That version seems more compliant with sklearn conventions, but doesn't seem to have a regularisation option. Regularisation is particularly interesting in that it allows for model overparameterisation (more features than training samples).