Blurs Behave Like Ensembles: Spatial Smoothings to Improve Accuracy, Uncertainty, and Robustness
Authors: Namuk Park, Songkuk Kim
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | By simply adding a few blur layers to the models, we empirically show that spatial smoothing improves accuracy, uncertainty estimation, and robustness of BNNs across a whole range of ensemble sizes. |
| Researcher Affiliation | Collaboration | 1NAVER AI Lab 2Yonsei University. Correspondence to: Namuk Park <namuk.park@navercorp.com>, Songkuk Kim <songkuk@yonsei.ac.kr>. |
| Pseudocode | Yes | Algorithm 1 Hessian max eigenvalue spectrum |
| Open Source Code | Yes | Code is available at https://github.com/ xxxnell/spatial-smoothing. |
| Open Datasets | Yes | We obtain the main experimental results with the Intel Xeon W-2123 Processor, 32GB memory, and a single Ge Force RTX 2080 Ti for CIFAR (Krizhevsky et al., 2009) and Cam Vid (Brostow et al., 2008). For Image Net (Russakovsky et al., 2015)... |
| Dataset Splits | No | The paper mentions using CIFAR, ImageNet, and CamVid datasets but does not explicitly provide the specific percentages or sample counts for training/validation/test splits, nor does it reference predefined splits with citations for these specific split details. |
| Hardware Specification | Yes | We obtain the main experimental results with the Intel Xeon W-2123 Processor, 32GB memory, and a single Ge Force RTX 2080 Ti for CIFAR (Krizhevsky et al., 2009) and Cam Vid (Brostow et al., 2008). For Image Net (Russakovsky et al., 2015), we use AMD Ryzen Threadripper 3960X 24Core Processor, 256GB memory, and four Ge Force RTX 2080 Ti. We conduct ablation studies with four Intel Intel Broadwell CPUs, 15GB memory, and a single NVIDIA T4. |
| Software Dependencies | No | The paper states "Models are implemented in Py Torch (Paszke et al., 2019)", but does not provide specific version numbers for PyTorch or any other software dependencies used for replication. |
| Experiment Setup | Yes | NNs are trained using categorical cross-entropy loss and SGD optimizer with initial learning rate of 0.1, momentum of 0.9, and weight decay of 5 10 4. We also use multistep learning rate scheduler with milestones at 60, 130, and 160, and gamma of 0.2 on CIFAR, and with milestones at 30, 60, and 80, and gamma of 0.2 on Image Net. We train NNs for 200 epochs with batch size of 128 on CIFAR, and for 90 epochs with batch size of 256 on Image Net. We use hyperparameters that minimizes NLL of Res Net: τ = 10, and MC dropout rate of 30% for CIFAR and 5% for Image Net. |