Partially Stochastic Infinitely Deep Bayesian Neural Networks
Authors: Sergio Calvo Ordoñez, Matthieu Meunier, Francesco Piatti, Yuantao Shi
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Lastly, empirical evaluations across multiple tasks show that our proposed architectures achieve better downstream task performance and uncertainty quantification than their counterparts while being significantly more efficient. In this section, we detail our Partially Stochastic Infinitely Deep BNNs performance for image classification, uncertainty quantification, and learning efficiency. We compare these results with fully stochastic, partially stochastic and deterministic baselines, and we perform ablation studies to understand the impact of key hyperparameters. |
| Researcher Affiliation | Academia | 1Oxford-Man Institute of Quantitative Finance, University of Oxford 2Mathematical Institute, University of Oxford 3Department of Mathematics, Imperial College London. |
| Pseudocode | No | The paper describes methods and processes through prose and mathematical formulations, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code can be found at https://github.com/Sergio20f/part_stoch_inf_deep |
| Open Datasets | Yes | For the MNIST (Le Cun et al., 2010) dataset, a single SDE-BNN block is utilised, whereas, for CIFAR-10 (Krizhevsky, 2009), we employ a multi-scale approach with several SDE-BNN blocks with the invertible downsampling from Dinh et al. (2016) in between. |
| Dataset Splits | No | The paper references the use of training, validation, and test accuracies/ECEs in figures and discussions, but it does not explicitly state the specific dataset split percentages or sample counts for these divisions. |
| Hardware Specification | Yes | Lastly, experiments were conducted on a single Nvidia RTX 3090 GPU. |
| Software Dependencies | No | The paper provides hyperparameter settings and mentions general activation functions and solvers, but it does not specify software dependencies with version numbers (e.g., Python, PyTorch, or CUDA versions). |
| Experiment Setup | Yes | Table 5. Hyperparameter settings for model evaluations in the classification tasks as presented in Table 1. Experiments were conducted on a single Nvidia RTX 3090 GPU within our compute clusters. Throughout the training, no scheduling was applied to the hyperparameters. |