From 1e3128dcadccda077a625e050812e0ecb30296f6 Mon Sep 17 00:00:00 2001 From: Steve Schmerler Date: Thu, 15 Jan 2026 14:49:00 +0100 Subject: [PATCH] First draft of Laplace model card --- docs/model_cards/laplace_approx.md | 27 +++++++++++++++++++++++++++ 1 file changed, 27 insertions(+) create mode 100644 docs/model_cards/laplace_approx.md diff --git a/docs/model_cards/laplace_approx.md b/docs/model_cards/laplace_approx.md new file mode 100644 index 0000000..8a10afe --- /dev/null +++ b/docs/model_cards/laplace_approx.md @@ -0,0 +1,27 @@ +# Laplace approximation + +## Abstract + +This is a post-processing method after NN training that constructs an approximate Gaussian posterior over model weights. Sampling from this posterior yields an "ensemble" which can be used to construct predictive uncertainty. For certain likelihoods (Gaussian in regression, non-Gaussian + further appproximations in classification), a closed form posterior predictive distribution is available, thus avoiding sampling. + + +## Easy to apply ⭐️⭐️⭐️★★ + +Easy when using exisiting libraries. Using the method involves and optimization procedure, similar to hyper-parameter tuning in a Gaussian Process. + +## Data Compatibility ⭐⭐️⭐️★★ + +Anything that the trained model ingests. + +## Task Compatibility ⭐️⭐️⭐️★★ + +Existing libraries: Classification or regression. See also https://lightning-uq-box.readthedocs.io/en/latest/#classification-of-uq-methods + +## Ease of integration ⭐️⭐️⭐️★★ + +Only for neural net models if you use existing libraries. +There are libraries for applying Laplace to PyTorch (https://github.com/aleximmer/Laplace) or JAX-based models (https://github.com/laplax-org/laplax). No model modifications are needed. In the PyTorch case, anything resembling a torch Module should work. The PyTorch library also has support for LLMs from huggingface. + +## Computationally Cheap ⭐️⭐️⭐️⭐️★ + +Given proper approximate methods (last layer, KFAC Hessian factorization), fitting Laplace and doing prediction requires roughly 10% the time of a typical train run.