Reparameterization invariance in approximate Bayesian inference

Authors: Hrittik Roy, Marco Miani, Carl Henrik Ek, Philipp Hennig, Marvin Pförtner, Lukas Tatzel, Søren Hauberg

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, our diffusion consistently improves posterior fit, suggesting that reparameterizations should be given more attention in Bayesian deep learning.
Researcher Affiliation Academia Hrittik Roy , Marco Miani Technical University of Denmark {hroy, mmia}@dtu.dk Carl Henrik Ek University of Cambridge, Karolinska Institutet che29@cam.ac.uk Philipp Hennig, Marvin Pförtner, Lukas Tatzel University of Tübingen, Tübingen AI Center {philipp.hennig, lukas.tatzel, marvin.pfoertner}@uni-tuebingen.de Søren Hauberg Technical University of Denmark sohau@dtu.dk
Pseudocode Yes Algorithm 1 Laplace diffusion
Open Source Code Yes Code: https://github.com/h-roy/geometric-laplace.
Open Datasets Yes We train a 44,000-parameter Le Net(Le Cun et al., 1989) on MNIST and FMNIST as well as a 270,000-parameter Res Net(He et al., 2016) on CIFAR-10(Krizhevsky et al., 2009).
Dataset Splits No The paper mentions 'held-out test set' but does not explicitly specify validation dataset splits or how they were derived for reproduction.
Hardware Specification Yes We run the sampling algorithm on H100 GPUs to run the high-order Lanczos decomposition.
Software Dependencies No The paper mentions 'Adam optimizer' and 'SGD' but does not specify software versions for libraries or frameworks like PyTorch, TensorFlow, or Python.
Experiment Setup Yes We train Le Net with Adam optimizer and a learning rate of 10 3. For the Re Net we use SGD with a learning rate of 0.1 with momentum and weight decay.