Sampling-based inference for large linear models, with application to linearised Laplace
Authors: Javier Antoran, Shreyas Padhy, Riccardo Barbano, Eric Nalisnick, David Janz, José Miguel Hernández-Lobato
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the strength of our inference technique in the context of the linearised Laplace procedure for image classification on CIFAR100 (100 classes 50k datapoints) using an 11M parameter Res Net-18. We also consider a high-resolution (251k pixel) tomographic reconstruction (regression) task with a 2M parameter U-Net. |
| Researcher Affiliation | Academia | Javier Antorán University of Cambridge Shreyas Padhy University of Cambridge Riccardo Barbano University College London Eric Nalisnick University of Amsterdam David Janz University of Alberta José Miguel Hernández-Lobato University of Cambridge |
| Pseudocode | Yes | Algorithm 1 summarises our method, Figure 1 shows an illustrative example, and full algorithmic detail is in Appendix F. |
| Open Source Code | Yes | An implementation of our method in JAX can be found here. Additional experimental results are provided in Appendix H and Appendix I. |
| Open Datasets | Yes | We demonstrate the strength of our inference technique in the context of the linearised Laplace procedure for image classification on CIFAR100 (100 classes 50k datapoints) using an 11M parameter Res Net-18. |
| Dataset Splits | Yes | The training set consists of n=60k observations and we employ 3 Le Net-style CNNs of increasing size: Le Net Small (d =14634), Le Net (d =29226) and Le Net Big (d =46024). |
| Hardware Specification | Yes | The latter is the largest model for which we can store the covariance matrix on an A100 GPU. |
| Software Dependencies | No | No specific version numbers for software dependencies were provided. The paper mentions 'JAX', 'GPy Torch', and 'Py Torch torchvision' but without their respective versions. |
| Experiment Setup | Yes | Appendix G provides full experimental details for all datasets and models used in our experiments. |