Summary of On Local Overfitting and Forgetting in Deep Neural Networks, by Uri Stern et al.
On Local Overfitting and Forgetting in Deep Neural Networks
by Uri Stern, Tomer Yaacoby, Daphna Weinshall
First submitted to arxiv on: 17 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 proposes a novel score to capture the forgetting rate of deep models on validation data, which quantifies local overfitting: a decline in performance confined to certain regions of the data space. The authors empirically show that local overfitting occurs regardless of traditional overfitting and provide a theoretical characterization of forgotten knowledge using deep over-parametrized linear models. They also devise an ensemble method to recover forgotten knowledge, which enhances model performance without adding inference costs when combined with self-distillation. Extensive evaluations demonstrate the efficacy of this method across multiple datasets, neural network architectures, and training protocols. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper tries to solve a puzzle: why do big neural networks often get better even though they’re more likely to overfit? They found that sometimes, overfitting happens in specific areas of the data, not just everywhere. To fix this, they created a new way to measure how well a model is doing on some validation data and found that it’s related to the knowledge the model forgot. Then, they developed an ensemble method that helps recover that forgotten knowledge without adding extra computation time. They tested their method with different datasets, neural networks, and training methods and showed that it works well. |
Keywords
» Artificial intelligence » Distillation » Inference » Neural network » Overfitting
