The fortnightly BRAG meeting will be held next Thursday 18th of April at 1 pm via Zoom/GP-Y801. In the upcoming week, we will have presentations by Ethan and Matt.
Zoom Link: https://qut.zoom.us/j/85658434714?pwd=a3ZMNTRRbXRmYkJPanRhdmJ6OVVBdz09
Password: 261088
Ethan’s Talk
Title: Tools of Bayesian Analysis for Explainable Deep Learning
Abstract: With deep learning providing unprecedented predictive capabilities for challenging perception tasks there is great interest in developing these models within a more principled probabilistic framework. Despite the interest in deploying a fully Bayesian approach to fitting these models, the increase in model size and computational complexity of modern networks quickly makes this infeasible. This talk will show how tools of Bayesian analysis can be applied to existing point-estimate networks within frameworks currently developed for explainable AI systems. We will show how adopting a probabilistic approach for explainable AI can allow for visualisation of features contributing to the strong predictive performance of existing deep-learning models whilst also highlighting features that give rise to predictive uncertainty.
Matt’s Talk
Title: Variance reduction in model choice using control variates
Abstract: Control variates (CV) can offer massive computational savings via post processing MCMC sampler output. This talk will discuss an application of CV for model choice based on a recent book chapter draft (with Dr Leah South) https://arxiv.org/abs/2402.07349. A snapshot of the results are shown below. The standard Monte Carlo approach was run for 10,000 iterations, two CV methods are applied one gives approximately state-of-the-art performance (= Rao Blackwellisation an alternative variance reduction method) and the other outperforms it achieving almost zero-variance performance.