We have just reposted our paper Bayesian Adversarial Privacy on arXiv to reflect the revision we wrote in the past months, to address the (quite sensible) comments from the reviewers. Interestingly the discussants of my Akaike lecture made similar points. The main changes are in explaining more clearly the nature of the combined loss, with Antoine coming up with the use of illuminating R-U map representations, in enlarging the references to other approaches, in stressing that Eve was an Alice’s construct rather than a genuine adversary, but still integrating the case of “multiple Eves”, in mellowing our criticisms of DP, and in expanding the conclusion with limitations and extensions subsections.
Bob then spoke about the latest version of NUTS, the within-orbit adaptive NUTS (WALNUTS) sampler, which adapts the step size at every leapfrog step in order to conserve the Hamiltonian and keep the path stable enough. The adaptation is facilitated by incorporating this step size as an extra parameter with an attached distribution, that the authors call Gibbs self tuning (GIST), for coupling tuning parameters and conditionally Gibbs-sampling them per iteration in Hamiltonian Monte Carlo. This has been done in the past, incl. in some of my papers (e.g., Andrieu & Robert, 2004), but I could not cite a particular reference during the seminar.
Further light reflections that came to mind during Bob’s talk:
