Tutorial on Functional Bayesian Deep Learning

Australasian Joint Conference on Artificial Intelligence (AJCAI), Melbourne, Australia, 25-29 Nov, 2024

Brief Description

Bayesian deep learning (BDL) is an emerging field that combines the strong function approximation power of deep learning with the uncertainty modelling capabilities of Bayesian inference. This synergy is poised to enhance model generalization and robustness, offering valuable uncertainty estimations for a range of safety-critical applications, including medical diagnostics, diabetes detection, autonomous driving, and civil aviation. Despite these advantages, the fusion introduces complexities to classical posterior inference in parameter space, such as nonmeaningful priors, intricate posteriors, and possible pathologies. This tutorial will delve into the driving forces, concepts, and methodologies underpinning BDL in function space, segueing into pivotal technological breakthroughs and their applications in machine learning tasks. To conclude, we will explore the prevailing hurdle faced by BDL.

Web Development

Outline

  1. Bayesian Deep Learning vs. Deterministic Deep Learning (50 mins)

    • Motivation (10 mins)

    • Uncertainty: What, Why, How (10 mins)

    • Bayesian deep learning: definition and category (10 mins)

    • Problems of weight-space BDL (10 mins)

    • break (10 mins)

  2. Function-space BDL vs. Weight-space BDL  (70 mins)

    • Functional maximum a posteriori (10 mins)

    • Functional variational inference (20 mins)

    • Functional Particle Optimization (10 mins)

    • Functional Markov Chain Monte Carlo (20 mins)

    • break (10 mins)

  3. Applications in Machine Learning Tasks (20 mins)

    • Continual Learning (10 mins)

    • Reinforcement Learning (10 mins)

  4. Challenge – Scalability (20 mins)

    • Current state-of-the-art (10 mins)

    • Efforts to improve efficiency (10 mins)

Venue and Time

13:30 ~ 17:00, (Monday) 25 November 2024

Room 03.01, Level 3, RMIT University Building 80, 445 Swanston Street, Melbourne, VIC, 3000.

Target Audience

The audiences are research scientists, postgraduate students, and deep learning practitioners who are already familiar with standard deep learning approaches and are interested in using Bayesian methods to enhance their understanding of uncertainty in neural network predictions.

Web Development

Speaker Information

Dr Junyu Xuan

Dr Xuan is an IEEE/ACM Senior Member, ISBA/BNP Life Member, ARC Discovery Early Career Researcher Award (DECRA) Fellow, and Senior Lecturer of Australia Artificial Intelligence Institute in the Faculty of Engineering and IT at the University of Technology Sydney (UTS). His research interests include Probabilistic Machine Learning, Bayesian Nonparametric Learning, Bayesian Deep Learning, Reinforcement Learning, Text Mining, Graph Neural Networks, etc. He has published over 60 papers in high-quality journals and conferences, including Artificial Intelligence Journal, Machine Learning Journal, IEEE TNNLS, IEEE TKDE, ACM Computing Surveys, ICDM, NIPS, AAAI, etc. He served as PC or Senior PC member for conferences, e.g. NIPS, ICML, UAI, ICLR, AABI, IJCAI, AAAI, EMNLP, etc.

John Doe

Contact

If you have any questions, please contact us at Junyu.Xuan@uts.edu.au.