Summary of Delve Into Base-novel Confusion: Redundancy Exploration For Few-shot Class-incremental Learning, by Haichen Zhou et al.
Delve into Base-Novel Confusion: Redundancy Exploration for Few-Shot Class-Incremental Learning
by Haichen Zhou, Yixiong Zou, Ruixuan Li, Yuhua Li, Kui Xiao
First submitted to arxiv on: 8 May 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
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 The paper proposes a novel approach to few-shot class-incremental learning (FSCIL), which aims to retain knowledge about base classes while acquiring new information from novel classes with limited samples. The existing methods for addressing catastrophic forgetting and overfitting by freezing the feature extractor during novel-class learning often cause confusion between base and novel classes, misclassifying novel-class samples into base classes. This paper delves into the phenomenon of this confusion, identifying it as a collision between the novel-class and base-class regions in the feature space caused by label-irrelevant redundancies within the base-class feature and pixel space. The authors propose a method called Redundancy Decoupling and Integration (RDI) to alleviate this collision, which decouples redundancies from base-class space to shrink the intra-base-class feature space and integrates them as a dummy class to enlarge the inter-base-class feature space. This approach achieves state-of-the-art performance on benchmark datasets such as CIFAR-100, miniImageNet, and CUB-200-2011. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper is about a way to learn new things from small amounts of data while keeping information we already know. The problem is that current methods for doing this often mix up the old information with the new. This paper figures out why this happens and proposes a solution called RDI (Redundancy Decoupling and Integration). It works by removing some extra information from what we already knew, which makes room for the new information to fit in. The result is better performance on several benchmark tests. |
Keywords
» Artificial intelligence » Few shot » Overfitting
