Summary of Interlude: Interactions Between Labeled and Unlabeled Data to Enhance Semi-supervised Learning, by Zhe Huang et al.
InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning
by Zhe Huang, Xiaowei Yu, Dajiang Zhu, Michael C. Hughes
First submitted to arxiv on: 15 Mar 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Medium Difficulty summary: Semi-supervised learning (SSL) aims to boost task performance by utilizing both labeled and unlabeled data. Traditional SSL image classification methods primarily focus on optimizing a loss that combines supervised classification objectives with regularization terms derived from unlabeled data. However, this formulation disregards the potential interaction between labeled and unlabeled images. To address this limitation, we introduce InterLUDE, a novel approach consisting of two parts: embedding fusion, which interpolates between labeled and unlabeled embeddings to enhance representation learning, and a new loss function grounded in consistency regularization, aiming to minimize discrepancies in model predictions between labeled and unlabeled inputs. Our method demonstrates clear benefits on standard closed-set SSL benchmarks and a medical SSL task with an uncurrated unlabeled set. Specifically, on the STL-10 dataset with only 40 labels, InterLUDE achieves an impressive error rate of 3.2%, outperforming previous methods with an error rate of 14.9%. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Low Difficulty summary: This paper is about a new way to improve machine learning using both labeled and unlabeled data. Traditional methods mix the two types of data without considering how they interact. The authors introduce InterLUDE, which has two main parts: one that combines information from labeled and unlabeled images to create better representations, and another that tries to make sure the model makes consistent predictions about labeled and unlabeled images. This new approach works well on several tests, including a medical task where it outperforms previous methods. |
Keywords
* Artificial intelligence * Classification * Embedding * Image classification * Loss function * Machine learning * Regularization * Representation learning * Semi supervised * Supervised
