On Uncertainty, Tempering, and Data Augmentation in Bayesian Classification
Authors: Sanyam Kapoor, Wesley J. Maddox, Pavel Izmailov, Andrew G. Wilson
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section we provide empirical support for the observations presented in Sections 4 and 5. First, in Section 6.1 we visualize the effect of data augmentation on the limiting distribution of SGMCMC using a Gaussian process regression model. Next, in Section 6.2 we report the results for BNNs on image classification problems. |
| Researcher Affiliation | Academia | Sanyam Kapoor New York University Wesley J. Maddox New York University Pavel Izmailov New York University Andrew Gordon Wilson New York University |
| Pseudocode | No | No pseudocode or clearly labeled algorithm block is present in the paper. |
| Open Source Code | Yes | The code to reproduce experiments is available at github.com/activatedgeek/understanding-bayesian-classification. |
| Open Datasets | Yes | For all experiments we use a Res Net-18 model [20] and the CIFAR-10 [35] and Tiny Imagenet [36] datasets. |
| Dataset Splits | No | The paper mentions using CIFAR-10 and Tiny Imagenet but does not explicitly state the dataset split percentages (e.g., train/validation/test splits) or specific sample counts for each split. |
| Hardware Specification | Yes | All experiments were performed on a single NVIDIA A100 GPU. |
| Software Dependencies | No | The paper mentions using PyTorch and Weights and Biases, but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | We use Adam with learning rate 5e-4 and a cosine annealing schedule to optimize the parameters of the model. We train for 100 epochs with batch size 128. For SGLD, we use the cyclical learning rate schedule, with an initial learning rate of 1e-2, a minimum learning rate of 1e-4, and a cycle length of 10 epochs. We collect 50 samples from the last 5 epochs of each cycle. |