Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Variational Inference on the Final-Layer Output of Neural Networks
Authors: Yadi Wei, Roni Khardon
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that VIFO provides a good tradeoff in terms of run time and uncertainty quantification, especially for out of distribution data. [...] 6 Experiments |
| Researcher Affiliation | Academia | Yadi Wei EMAIL Luddy School of Informatics, Computing, and Engineering Indiana University Roni Khardon EMAIL Luddy School of Informatics, Computing, and Engineering Indiana University |
| Pseudocode | No | The paper describes mathematical derivations and methods in prose and equations (e.g., Eq. 1, Eq. 3, Eq. 4, etc.) but does not include any explicitly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Our code is available on https://github.com/weiyadi/VIFO. |
| Open Datasets | Yes | For our main experiments, we pick four large datasets, CIFAR10, CIFAR100, SVHN, STL10, together with two types of neural networks, Alex Net (Krizhevsky et al., 2012) and Pre Res Net20 (He et al., 2016). |
| Dataset Splits | Yes | To generate Fig. 1 and Fig. 2, we generate 100 training data points y = 2 sin x + 0.1ϵ, ϵ N(0, 1), where xtrain [ 3 4π] and xtest [ π, π]. [...] For our main experiments, we pick four large datasets, CIFAR10, CIFAR100, SVHN, STL10. |
| Hardware Specification | No | Some of the experiments in this paper were run on the Big Red computing system at Indiana University, supported in part by Lilly Endowment, Inc., through its support for the Indiana University Pervasive Technology Institute. |
| Software Dependencies | No | For all methods other than SGD, SWA and SWAG, we use the Adam optimizer with learning rate 0.001. |
| Experiment Setup | Yes | Number of training epochs: We train all methods in 500 epochs. Learning rate: For all methods other than SGD, SWA and SWAG, we use the Adam optimizer with learning rate 0.001. [...] For both VI and VIFO, the regularization parameter η is fixed at 0.1. |