I will discuss methods to turn manifold learning methods into nonlinear collective variables usable in accelerated molecular dynamics simulations. These methods include the systematic smooth parametrization of a slow manifold, the construction of a differentiable closest-point map, and the consideration of an atlas of local charts to describe systems with configuration spaces of complex geometry or topology. I will also discuss research ideas that we did not bring to fruition. (Joint work with Behrooz Hashemian)