Summary of The Effect Of Intrinsic Dataset Properties on Generalization: Unraveling Learning Differences Between Natural and Medical Images, by Nicholas Konz et al.
The Effect of Intrinsic Dataset Properties on Generalization: Unraveling Learning Differences Between Natural and Medical Images
by Nicholas Konz, Maciej A. Mazurowski
First submitted to arxiv on: 16 Jan 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG); Image and Video Processing (eess.IV); Machine Learning (stat.ML)
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 investigates the discrepancies in how neural networks learn from different imaging domains, particularly between medical and natural images. It reveals that the generalization error of a trained network increases with the intrinsic dimension of its training set, but this relationship varies significantly between the two domains. The study establishes a generalization scaling law and attributes the discrepancy to the higher label sharpness of medical imaging datasets. Additionally, it shows that measuring label sharpness is negatively correlated with adversarial robustness, making models for medical images more vulnerable to attacks. The paper also extends its formalism to learned representation intrinsic dimension and demonstrates an upper bound between the two metrics. Experiments with six models and eleven datasets support these findings, providing insights into the influence of dataset properties on generalization, representation learning, and robustness in deep neural networks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at how well artificial intelligence (AI) learns from different types of images. It’s like trying to learn a new language – if you only practice with simple words, you might not do well when faced with complex sentences. The study found that AI does better when learning from natural images than medical images, and it thinks this is because medical images are harder to understand (like complex sentences). This could make AI more vulnerable to being tricked or fooled. The paper also shows that how well AI learns depends on the quality of the training data. |
Keywords
* Artificial intelligence * Generalization * Representation learning
