//! Learning algorithms, written to be read rather than to be fast.
//!
//! Every method here has a closed-form or exactly-checkable property
//! attached to it, because that is what makes a learning algorithm
//! testable at all. A network that trains to a plausible loss is not
//! evidence of anything -- gradient descent will happily reduce the loss
//! of a model whose gradients are wrong, just more slowly. What settles
//! it is comparing the analytic gradient against a finite difference,
//! comparing a linear model fitted by descent against the normal
//! equations, or checking that a clustering agrees with itself under a
//! relabelling.