Summary of Training Data Attribution Via Approximate Unrolled Differentiation, by Juhan Bae et al.
Training Data Attribution via Approximate Unrolled Differentiation
by Juhan Bae, Wu Lin, Jonathan Lorraine, Roger Grosse
First submitted to arxiv on: 20 May 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 The paper proposes a new method called Source for estimating the effect of removing individual data points from a training set on a model’s behavior. This is achieved by combining benefits from both implicit differentiation and unrolling-based approaches. The method, an approximate unrolling-based TDA approach, uses an influence-function-like formula to estimate how a model would behave if certain data points were removed. This approach is computationally efficient compared to unrolling-based methods and suitable for cases where implicit-differentiation-based approaches struggle, such as non-converged models or multi-stage training pipelines. Empirically, Source outperforms existing TDA techniques in counterfactual prediction. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Source is a new method that helps us understand how changing the data used to train a model would affect its behavior. It combines two different ways of doing this: implicit differentiation and unrolling. The result is an efficient way to estimate the effect of removing individual data points from a training set on a model’s behavior. This is helpful in situations where the model isn’t fully trained or when we’re using a multi-stage training pipeline. |
