Naively, one might just try to use variational inference to train a BNN with a GP prior, but it has been pointed out that this leads to some mathematical issues (infinite KL, etc.): https://arxiv.org/abs/2011.09421
Comments
Log in with your Bluesky account to leave a comment
💡 In our work, we propose to use the Laplace approximation in function space! This is mathematically principled (after a bit of measure theory) and can be efficiently implemented using matrix-free linear algebra 🚀
🌊 We show that this can improve performance over standard Laplace with weight-space priors in real-world scientific tasks, such as this ocean current modeling problem
Comments
📅 Thursday, 4:30 PM
📍Poster #3907
Or read the paper here: https://arxiv.org/abs/2407.13711