AEST via online
The SSA Bayes Branch Proudly Presents this Webinar.
It is well known that the posterior distribution over neural network weights can be approximated by neither a Gaussian nor a Gaussian mixture distribution. Rather, as established in singular learning theory, the posterior distribution over the parameters of a singular model is, asymptotically, a mixture of standard forms. Namely, there exists a partition of the parameter space such that in each local parameter set, the average log likelihood ratio can be made "normal crossing" via an algebraic geometrical transform known as a resolution map. We leverage this under-appreciated result to propose a generalized gamma mean-field variational family which can recover the leading term of the (normalized) log evidence. Affine coupling layers are employed to learn the unknown resolution map, effectively rendering the proposed methodology a normalizing flow with the generalized gamma as the source distribution, rather than the multivariate Gaussian typically employed.
Presenter: Dr Susan Wei (DECRA fellow, University of Melbourne)
Susan Wei is a lecturer in the School of Mathematics and Statistics at the University of Melbourne. She currently holds a Discovery Early Career Researcher Award (DECRA) from the Australian Research Council (ARC). Her research interests include statistics, machine learning, and deep learning.
The event is free but you do need to register. To register click here.
@StatSocAus
