First-order methods and convex toolkit for training models — from gradient descent to Adam, with enough convexity to know what your optimizer can and cannot guarantee.