Hacker News Reader

Comment by hyperbovine | Hacker News Reader

Were you optimizing over the knots as well? Otherwise I can't see why this would be disallowed using either forward or reverse-mode AD. An infinitesimal perturbation of beta will not cause t * exp(-beta^T x) to cross a knot, so the whole thing is smooth. (And, with B-splines the derivatives are continuous from piece to piece anyways.) But in general I agree--a good spline implementation I something I miss the most when moving from scipy.interpolate to jax.scipy. Given that the SciPy implementation is mostly F77 code written before I was born, I do not see this situation resolving itself anytime soon.

View on HN

A short answer: this requires a basis matrix for the splines rather than interpolation.

A longer answer: the splines require a basis matrix B(t), do that g(S_0(t))= B(t) gamma for a vector of parameters gamma and some transformation g of survival. A classical choice would be to use M-splines and I-splines with g(S)=-log(S) and a penalised likelihood, with the constraint that the gammas should be increasing. In R, this would use the splines2 package, while in Julia, one could use the splines2.jl package (disclaimer: which I maintain). The computational challenge is that the basis matrices need to be re-evaluated for changes in the coefficients for the covariates (that is, the betas).

It's not about smoothness, it's about how to JIT the gradient function. ML libs don't generally do interpolation, partly because it's tricky to vectorize (you have to search for which segment to use for each element) and partly because most ML practioners don't need it. What I've done in my code is use all the vertices for all the elements, but with weights that are mostly zero. It's pretty fast on GPU because I don't use that many vertices.

"Interpolation" here I think effectively boils down to np.searchsorted() or its equivalent, which is implemented in all the major ML libs (jittable and backpropable).

Reply on news.ycombinator.com