Bayesian Low-Rank Learning (Bella): A Practical Approach to Bayesian Neural Networks
Authors: Bao Gia Doan, Afshar Shamsi, Xiao-Yu Guo, Arash Mohammadi, Hamid Alinejad-Rokny, Dino Sejdinovic, Damien Teney, Damith C. Ranasinghe, Ehsan Abbasnejad
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive empirical evaluation in large-scale tasks such as Image Net, CAMELYON17, Domain Net, VQA with CLIP, LLa VA demonstrate the effectiveness and versatility of Bella in building highly scalable and practical Bayesian deep models for real-world applications. |
| Researcher Affiliation | Academia | 1The University of Adelaide 2Concordia University 3Idiap Research Institute 4University of New South Wales EMAIL; EMAIL; EMAIL; EMAIL |
| Pseudocode | No | The paper describes mathematical formulations for SVGD and Bella updates but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures in the main text. |
| Open Source Code | Yes | Code https://bnn-bella.github.io/BNN-Bella/ |
| Open Datasets | Yes | Datasets. In this research, we have employed a variety of datasets, each selected for their relevance and contribution, they include CIFAR-10, CIFAR-100 (Krizhevsky, Hinton et al. 2009), CIFAR-10-C (Hendrycks and Dietterich 2019), STL-10 (Coates, Ng, and Lee 2011), CAMELYON17 (Bandi et al. 2018), Image Net (Russakovsky et al. 2015), and Domain Net (Peng et al. 2019). We also consider VQA v2 dataset utilized for Visual Question Answering (VQA). |
| Dataset Splits | No | The paper mentions using well-known datasets like CIFAR-10, CAMELYON17, ImageNet, and VQA v2, implying the use of standard splits. However, it does not explicitly provide the specific training/test/validation split percentages or sample counts used for the experiments in the main text. It defers some details to the Appendix. |
| Hardware Specification | Yes | Notably, with SVGD Baseline Models, we can only train up to n=40 particles on a A6000 48 GB GPU, while we can increase to more than 100 parameter particles with our Bella method with negligible increase of GPU consumption. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as programming languages, libraries, or frameworks (e.g., Python, PyTorch, CUDA versions) used for implementation. |
| Experiment Setup | No | The paper states: 'Detailed hyper-parameters are in the Appendix.' and 'Further details about the dataset, model and metrics are deferred to the Appendix.' This indicates that specific experimental setup details, including hyperparameters, are not provided in the main text. |