InPL: Pseudo-labeling the Inliers First for Imbalanced Semi-supervised Learning
Authors: Zhuoran Yu, Yin Li, Yong Jae Lee
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate that our energy-based pseudo-labeling method, In PL, albeit conceptually simple, significantly outperforms confidence-based methods on imbalanced SSL benchmarks. For example, it produces around 3% absolute accuracy improvement on CIFAR10-LT. When combined with state-of-the-art long-tailed SSL methods, further improvements are attained. In particular, in one of the most challenging scenarios, In PL achieves a 6.9% accuracy improvement over the best competitor. To evaluate the proposed In PL, we integrate it into the classic Fix Match (Sohn et al., 2020) framework and the recent state-of-the-art imbalanced SSL framework ABC (Lee et al., 2021) by replacing their vanilla confidence-based pseudo-labeling with our energy-based pseudo-labeling. In this section, we first compare our energy-based pseudo-labeling approach to confidence-based pseudo-labeling approaches that build upon Fix Match (Sohn et al., 2020) and ABC (Lee et al., 2021) framework and compare to state-of-the-art imbalanced SSL approaches. We evaluate our method on both standard imbalanced SSL benchmarks and the large-scale Image Net dataset. |
| Researcher Affiliation | Academia | Zhuoran Yu Yin Li Yong Jae Lee University of Wisconsin-Madison {zhuoran.yu, yin.li}@wisc.edu yongjaelee@cs.wisc.edu |
| Pseudocode | No | The paper describes steps for its method in paragraph text but does not include a structured pseudocode block or algorithm figure. |
| Open Source Code | No | The paper does not provide an unambiguous statement about releasing source code for the described methodology or a direct link to a code repository. |
| Open Datasets | Yes | We evaluate on CIFAR10-LT and CIFAR100-LT (Krizhevsky & Hinton, 2009) and CIFAR10 (Krizhevsky & Hinton, 2009), SVHN (Netzer et al., 2011), and STL-10 (Coates et al., 2011). |
| Dataset Splits | Yes | For CIFAR10-LT and CIFAR100-LT, we use 10% and 30% data as labeled sets, respectively. (Table 1 caption) We use 20% labeled data for CIFAR10-LT and 40% labeled data for CIFAR100-LT. (Table 2 caption) The median value in last 20 evaluations of each single run is recorded and the results are reported as the mean of recorded values over three different runs. (Appendix A.2) |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. It only mentions general 'limited computation'. |
| Software Dependencies | No | The paper mentions optimizers (Adam, SGD) and network architectures (Wide Res Net-28-2) but does not provide specific version numbers for these or other software dependencies like programming languages or deep learning frameworks. |
| Experiment Setup | Yes | Specifically, for all the results, we use Adam (Kingma & Ba, 2014) as the optimizer with a constant learning rate 0.002 and train models with 25000 iterations. Exponential moving average of momentum 0.999 is used to maintain an ensemble network for evaluation. The strong augmentation transforms include Rand Augment and Cut Out, which is consistent with the standard practice in SSL. Batch size of both labeled set and unlabeled set is set to 64 and two strongly-augmented views of each unlabeled sample are included in training. The base learning rate is set to 0.03 and evaluation is conducted every 500 iterations with total number of training iterations being 250000. We use Wide Res Net-28-2 (Zagoruyko & Komodakis, 2016) for CIFAR 10-LT and WRN-28-8 for CIFAR 100-LT. |