Summary of Sea: Semantic Adversarial Augmentation For Last Layer Features From Unsupervised Representation Learning, by Qi Qian et al.
SeA: Semantic Adversarial Augmentation for Last Layer Features from Unsupervised Representation Learning
by Qi Qian, Yuanhong Xu, Juhua Hu
First submitted to arxiv on: 23 Aug 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 This paper investigates the potential of fixed deep features from pre-trained models for various downstream classification tasks. Researchers found that these features outperform traditional hand-crafted features. However, they also noted that exploring diverse augmentations in the original input space is more effective than trying to optimize them in the feature space. To address this challenge, the authors propose a novel semantic adversarial augmentation (SeA) approach. SeA projects the adversarial direction implied by the gradient onto a subspace spanned by other examples to preserve semantic information. Experiments were conducted on 11 benchmark tasks using four popular pre-trained models, demonstrating an average improvement of 2% compared to traditional deep features. Notably, SeA showed comparable performance to fine-tuning in six out of 11 tasks, highlighting its efficiency and effectiveness. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research looks at how well fixed features from pre-trained artificial intelligence models can be used for different classification tasks. The study found that these features work better than traditional ones. However, they also discovered that trying to improve these features is harder than just using them as they are. To make it easier, the researchers created a new way called SeA (Semantic Adversarial Augmentation). It helps keep the important information in the features while changing them slightly. The team tested SeA on 11 different tasks and found that it works almost as well as making the model learn everything from scratch. |
Keywords
» Artificial intelligence » Classification » Fine tuning
