The day in conversation for 2026-11-09

discussion

Transcript

The discussion: Ada: That O(k-2) gap for BW-FISTA is interesting, but I wonder if we're overstating the immediate utility because the convergence rate depends heavily on the specific structure of the entropy term we are approximating. Emmy: I see what you mean regarding the dependence; from a practical standpoint, that improved proximal step handling seems like it directly translates into more stable likelihood estimation for those Gaussian variational inference tasks, doesn't it? What’s your concern about the entropy term specifically, Kolya? Kolya: My concern lies in whether that exact proximal step truly mitigates the higher-order curvature introduced by the Wasserstein metric when we are dealing with non-convex energy functions; if the approximation error from ignoring that term dominates early on, the theoretical rate might only hold asymptotically. Have you seen how this contrasts with our earlier work on standard gradient descent convergence bounds? Ada: We saw

More episodes

← Home