Deep Active Learning by Leveraging Training Dynamics
Authors: Haonan Wang, Wei Huang, Ziwei Wu, Hanghang Tong, Andrew J Margenot, Jingrui He
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results show that dynamic AL not only outperforms the other baselines consistently but also scales well on large deep learning models. Regarding experiments, we have empirically verified our theory by conducting extensive experiments on three datasets, CIFAR10 [23], SVHN [24], and Caltech101 [25] using three types of network structures: vanilla CNN, Res Net [26], and VGG [27]. |
| Researcher Affiliation | Academia | 1University of Illinois Urbana-Champaign 2University of New South Wales |
| Pseudocode | Yes | Algorithm 1 Deep Active Learning by Leveraging Training Dynamics |
| Open Source Code | No | The paper does not provide an explicit statement about the release of its source code or a link to a repository. |
| Open Datasets | Yes | We evaluate all the methods on three benchmark data sets, namely, CIFAR10 [23], SVHN [24], and Caltech101 [25]. |
| Dataset Splits | No | The paper describes initial labeled sets and test sets but does not explicitly mention a validation set split for their experiments. |
| Hardware Specification | No | The paper does not specify the hardware used for running the experiments (e.g., CPU, GPU models, memory). |
| Software Dependencies | No | The paper mentions PyTorch in the references (e.g., [55] Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019.) but does not specify the version numbers for PyTorch or any other software dependencies used in their experiments. |
| Experiment Setup | Yes | We consider three neural network architectures: vanilla CNN, Res Net18 [26], and VGG11 [27]. For each model, we keep the hyper-parameters used in their official implementations. More information about the implementation is in Appendix C.1. ... batch size b varying from {250, 500, 1000}. ... initial set size M = 500 for all those three data sets. ... Train f( ; θ) on S with cross-entropy loss until convergence. ... re-initialization is not used after each query. |