[ML Seminar] IDA ML seminar - Samuel Matthiesen - VIKING: Deep variational inference with stochastic projections

[Louis Ohl]

Welcome to an IDA Machine Learning Seminar on Wednesday, February 11 at 13:30 In Alan Turing

VIKING: Deep variational inference with stochastic projections

Samuel Matthiesen<https://samuel.nihil.ws/>, Postdoc at DTU Compute

Abstract: Variational mean field approximations tend to struggle with contemporary overparameterised deep neural networks. Where a Bayesian treatment is usually associated with high-quality predictions and uncertainties, the practical reality has been the opposite, with unstable training, poor predictive power, and subpar calibration. Building upon recent work on reparameterisations of neural networks, we propose a simple variational family that considers two independent linear subspaces of the parameter space. These represent functional changes inside and outside the support of training data. This allows us to build a fully-correlated approximate posterior reflecting the overparameterisation that tunes easy-to-interpret hyperparameters. We develop scalable numerical routines that maximize the associated evidence lower bound (ELBO) and sample from the approximate posterior. Our results show that approximate Bayesian inference applied to deep neural networks is far from a lost cause when constructing inference mechanisms that reflect the geometry of reparametrisations.

Location: Alan Turing, https://link.mazemap.com/RDyxFvXH

---------------
The list of future seminars in the series is available at:
https://stima.gitlab-pages.liu.se/ml-seminars/
You can subscribe to the seminar series' calendar using this ics link: https://outlook.office365.com/owa/calendar/4d811ae47ce446f58d11a7c2f50a7ed8@ad.liu.se/0f5253d7bc7841248c71eb4c28eb2d668927992292494627279/calendar.ics

--
Louis Ohl
Postdoc
louis.ohl at liu.se

Linköping University
IDA / STIMA
B-Huset, 2A:445
Webpage<https://oshillou.github.io/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.liu.se/pipermail/ml-seminars/attachments/20260128/0c439919/attachment.htm>