Lightweight uncertainty modelling using function space particle optimization
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: uncertainty quantification, bayesian inference, function space particle optimization
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Deep ensembles have shown remarkable empirical success in quantifying uncertainty, albeit at considerable computational cost and memory footprint. Meanwhile, deterministic single-network uncertainty methods have proven as computationally effective alternatives, providing uncertainty estimates based on distributions of latent representations. While those methods are successful at out-of-domain detection, they exhibit poor calibration under distribution shifts. In this work, we propose a method that provides calibrated uncertainty by utilizing particle-based variational inference in function space. Rather than using full deep ensembles to represent particles in function space, we propose a single multi-headed neural network that is regularized to preserve bi-Lipschitz conditions. Sharing a joint latent representation enables a reduction in computational requirements, while prediction diversity is maintained by the multiple heads. We achieve competitive results in disentangling aleatoric and epistemic uncertainty for active learning, detecting out-of-domain data, and providing calibrated uncertainty estimates under distribution shifts while significantly reducing compute and memory requirements.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7470
Loading