A Neural Pre-Conditioning Active Learning Algorithm to Reduce Label Complexity
Authors: Seo Taek Kong, Soomin Jeon, Dongbin Na, Jaewon Lee, Hong-Seok Lee, Kyu-Hwan Jung
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Section 5 presents experiments that show how the proposed algorithm enhances downstream SSL performance, and highlights how some AL algorithms are not as effective in our proposed setting. Tables 1 and 2 show the accuracy of AL algorithms when trained on CIFAR-10 and CIFAR-100, respectively. |
| Researcher Affiliation | Collaboration | Seo Taek Kong1, , Soomin Jeon2 Dongbin Na3 Jaewon Lee3 Hong-Seok Lee3 Kyu-Hwan Jung4, , 1University of Illinois, Urbana-Champaign 2Dong-A University 3VUNO Inc. 4Sungkyunkwan University |
| Pseudocode | Yes | Algorithm 1 Neural Pre-Conditioning (batch-mode solution to (3)) Inputs: Unlabeled pool XU, acquisition size Q. Output: New pool X u to be labeled. for i = 1, , m = O (NU) do X (i) u Q unlabeled instances randomly sampled from XU. v(X (i) u ) λmin XL X (i) u using the network s Gram matrix. end for Return X u arg max (v) |
| Open Source Code | Yes | 3. If you ran experiments... (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] |
| Open Datasets | Yes | Tables 1 and 2 show the accuracy of AL algorithms when trained on CIFAR-10 and CIFAR-100, respectively. |
| Dataset Splits | No | The paper specifies random sampling and query sizes for CIFAR-10 and CIFAR-100 and mentions training from scratch, but it does not provide explicit training, validation, or test dataset splits (e.g., percentages or counts) or refer to standard predefined splits for these datasets. |
| Hardware Specification | No | 3. If you ran experiments... (d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No] Computational requirements follow standard semi-supervised learning settings. |
| Software Dependencies | No | The paper mentions specific algorithms and architectures like "Fix Match (Sohn et al., 2020)" and "WRN-28-2 architecture (Zagoruyko & Komodakis, 2016)", but it does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We adopt all SSL-related configurations from (Oliver et al., 2018) and use the WRN-28-2 architecture (Zagoruyko & Komodakis, 2016) for all experiments. At the first acquisition step, we randomly sampled 1 image per class and used the model that attained median performance across 5 trials. Subsequent acquisitions were performed with query size Q = 20 for CIFAR-10 and Q = 200 for CIFAR-100. All performances are averaged over 3 trials. Following most AL setups, we train classifiers from scratch after each acquisition step. |