Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning
Authors: Jaehyung Kim, Youngbum Hur, Sejun Park, Eunho Yang, Sung Ju Hwang, Jinwoo Shin
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate our algorithm on various scenarios for imbalanced semi-supervised learning in classification problems. We first describe the experimental setups in Section 4.1. In Section 4.2, we present empirical evaluations on DARP and other baseline algorithms under various setups. In Section 4.3, we present detailed analysis on DARP. Table 1 summarizes the performance of baseline algorithms with/without DARP for learning CIFAR-10. |
| Researcher Affiliation | Collaboration | Jaehyung Kim1, Youngbum Hur2, Sejun Park1, Eunho Yang1,3, Sung Ju Hwang1,3, Jinwoo Shin1 1Korea Advanced Institute of Science and Technology (KAIST) 2Samsung Advanced Institute of Technology 3AItrics |
| Pseudocode | Yes | Algorithm 1 Dual Coordinate Ascent: Coordinate ascent algorithm for dual of (1) and Algorithm 2 DARP: Distribution aligning refinery of pseudo-label |
| Open Source Code | No | The paper does not provide an explicit statement about open-sourcing the code for the described methodology or a link to a code repository. |
| Open Datasets | Yes | We consider synthetically long-tailed variants of CIFAR-10, CIFAR-100 [23], and STL-10 [11] in order to evaluate our algorithm under various levels of imbalance. Results on real-world dataset, SUN-397 [40], are also given in Section C of the supplementary material. |
| Dataset Splits | No | The paper mentions training data and test data, and a specific split for estimating confusion matrices for DARP ("We split the labeled dataset as Dlabeled = Dest Dtrain where Dest Dtrain = ."), but it does not specify a distinct validation set split for model selection during the main training process of the experimental models. It states "We evaluate the model on the test dataset for every 500 iterations and report the average test accuracy of the last 20 evaluations following [5]." |
| Hardware Specification | No | The paper does not explicitly state any specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | All experiments are conducted with Wide Res Net-28-2 [30] and it is trained with batch size 64 for 2.5 105 training iterations. For all algorithms, we evaluate the model on the test dataset for every 500 iterations and report the average test accuracy of the last 20 evaluations following [5]. We apply the DARP procedure for every 10 iterations with fixed hyper-parameters δ = 2 and T = 10, which is empirically enough for the convergence of DARP. Since pseudo-labels are not accurate at the early stage of training, we are not using DARP until the first 40% of iterations. |