Meta-Neighborhoods
Authors: Siyuan Shan, Yang Li, Junier B. Oliva
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conduct experiments for both classification and regression tasks. To demonstrate the benefits of making predictions in local neighborhoods, we compare to the vanilla model where the same network architecture is used but without the learnable dictionary, and MAXL [24] for classification task. |
| Researcher Affiliation | Academia | Siyuan Shan Department of Computer Science University of North Carolina at Chapel Hill siyuanshan@cs.unc.edu Yang Li Department of Computer Science University of North Carolina at Chapel Hill yangli95@cs.unc.edu Junier B. Oliva Department of Computer Science University of North Carolina at Chapel Hill joliva@cs.unc.edu |
| Pseudocode | Yes | The pseudocode of our training algorithm is given in Algorithm 1. |
| Open Source Code | Yes | 1The code is available at https://github.com/lupalab/Meta-Neighborhoods |
| Open Datasets | Yes | In this section, we evaluate 9 datasets with different complexities and sizes: MNIST [21], MNIST-M[7], PACS[22], SVHN [9], CIFAR-10 [19], CIFAR-100 [19], CINIC-10 [3], Tiny-Image Net [31] and Image Net [4]. and We use five publicly available datasets with various sizes from UCI Machine Learning Repository [5]. |
| Dataset Splits | Yes | For MNIST, MNIST-M, SVHN, CIFAR-10, CIFAR-100, CINIC-10, Tiny-Image Net and Image Net, we follow standard practice to use official train/test splits. We do not use a separate validation set, instead, we report test results at the end of training after the model converges. and 5-fold cross-validation is used to report the results in Table 2. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory specifications, or cloud instance types) used to run the experiments. |
| Software Dependencies | No | The paper mentions optimizers like Adam [18] and SGD, but does not provide specific version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | We use 100 induced neighbors. The labels of neighbors (values in the dictionary) are fixed after random initialization for illustration purposes. The 2D locations of neighbors (keys in the dictionary) are updated with the model. We use negative Euclidean distance as the similarity metric in (5) and set T to 0.1. and For all image datasets, we use Adam [18] optimizer with a learning rate of 1e-3, weight decay 1e-4, and batch size of 128. For PACS dataset, we apply standard data augmentation including random cropping and horizontal flipping. For other datasets, we use standard data augmentation except random cropping. For regression tasks, we use Adam [18] optimizer with a learning rate of 1e-4, weight decay 1e-4, and batch size of 64. |