Training diverse ensembles for OOD object detection

Last updated on Oct 26, 2020

The aim of this project was to investigate whether training a Bayesian Neural Network/Neural Network ensemble using Stein Variational Gradient Descent (instead of Bayes By Backprop or MC-Dropout) could lead to a more expressive and diverse ensemble, which more accurately reflects (or atleast underestimates to a relatively lesser extent) the uncertainty of the model. More details coming soon!

Deep Learning Autonomous Driving Variational Inference
Gunshi Gupta
Gunshi Gupta
Deep Learning Researcher

My research interests include Meta-Learning, Bayesian and Continual Deep Learning, Robotics.

Related