University of Cambridge > Talks.cam > Machine Learning @ CUED > Two Approximate Sampling Methods for Bayesian Deep Learning

  Browse A-Z

Log in

University Account External (via Google)

🔐 Log In

Information on

Subscribing to talksdetails Finding a talkdetails Adding a talkdetails Disseminating talksdetails Help and Documentationdetails

ℹ️ Information

Two Approximate Sampling Methods for Bayesian Deep Learning

Download to your calendar using vCal

If you have a question about this talk, please contact Adrià Garriga Alonso .

Deep neural networks have become popular for many tasks, especially object classification, in computer vision and machine learning. However, these classes of models are known to be have poor uncertainty representations – e.g. they do not know what they do not know. To address this challenge, we propose two Bayesian approaches to approximate the posterior distribution of the models’ parameters. The first, termed stochastic weight averaging Gaussian (SWAG), fits a Gaussian approximation around the iterates of the stochastic gradient descent trajectory from standard training of DNNs. The second, subspace inference, instead reduces the high dimensionality of DNNs to very low dimensions, before performing Bayesian model averaging in that low dimensional subspace. Both methods draw on existing theory and are demonstrated to have strong empirical results on both regression and classification, scaling to even ImageNet-sized datasets.

This talk is part of the Machine Learning @ CUED series.

This talk is included in these lists:

Note that ex-directory lists are not shown.

Other lists

Online - to be confirmed BAS Cambridge Seminar Room 1 and Zoom Cellular Genetic Disease Seminar

Other talks

Market Failure, Organ Shortage, Pandemic and Climate Change: The Architecture of Cooperation Tea and Coffee Break Organiser's Welcome Dwarf Galaxies in the Local Universe Coming soon The physics of decision: how competitive systems compute, route, and remember
 

© 2006-2025 Talks.cam, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity