Hyper Evidential Deep Learning to Quantify Composite Classification Uncertainty
Authors: Changbin Li, Kangshuo Li, Yuzhe Ou, Lance M. Kaplan, Audun Jøsang, Jin-Hee Cho, DONG HYUN JEONG, Feng Chen
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our results demonstrate that HENN outperforms its state-of-the-art counterparts based on four image datasets. The code and datasets are available at: https://github.com/ Hugo101/Hyper Evidential NN. |
| Researcher Affiliation | Collaboration | Department of Computer Science, The University of Texas at Dallas1, US Army Research Laboratory2, University of Oslo3, Virginia Tech4, University of the District of Columbia5 |
| Pseudocode | Yes | The pseudocode is shown in App.(Algo. 1). |
| Open Source Code | Yes | The code and datasets are available at: https://github.com/ Hugo101/Hyper Evidential NN. |
| Open Datasets | Yes | Tiny Image Net (Fei-Fei et al., 2015), Living17 (Santurkar et al., 2021), Nonliving26 (Santurkar et al., 2021), and CIFAR100 (Krizhevsky & Hinton, 2009) are used in the experiments. |
| Dataset Splits | Yes | We split the original training set into a training and a validation set according to the ratio 9:1. Therefore, the number of images per class will be: 450/50/50 for training/validation/test set for CIFAR100, similarly for other datasets. |
| Hardware Specification | No | The paper mentions using pre-trained model backbones like Efficient Net-b3, Res Net50, and VGG16, but it does not specify the actual hardware (e.g., GPU models, CPU types, or cloud computing instances) used to run the experiments. |
| Software Dependencies | No | The paper mentions using PyTorch for the Gaussian Blurring operation and Adam as the optimizer. However, it does not provide specific version numbers for these software components, which is necessary for reproducibility. |
| Experiment Setup | Yes | For our method and all other baselines, we adopt Adam (Kingma & Ba, 2014) as optimizer with parameters β1 = 0.9, β2 = 0.999, weight decay is 0, ϵ = 1e 8 provided in (Kingma & Ba, 2014). The number of epochs for all experiments is set to 100. Other hyperparameters used in this paper mainly are learning rate and weight of entropy regularizer. Grid search is leveraged to determine the best hyperparameters based on a held-out validation set for each specific experiment. Specifically, (1) DNN. the learning rate is chosen from {1e-5, 1e-4, 1e-3}; the cutoff is chosen from {0, 0.05, ..., 0.5}. (6) HENN. the learning rate is chosen from {1e-5, 1e-4, 1e-3} and the weight of regularizer λ is chosen from {1, 1e-1, 1e-2, 1e-3, 1e-4, 1e-5}. |