TL;DR
This paper introduces "Unlearning Shapley," a novel framework for efficiently valuing data contributions in large AI models. By leveraging machine unlearning, it bypasses the prohibitive computational costs of traditional methods, enabling both full and partial data valuation, which is crucial for modern data markets.
What Problem Does This Paper Address?
The rapid proliferation of large AI models has created an urgent need for effective data valuation methods. These methods are essential for fairly distributing profits in data markets, as they quantify how much each data provider contributes to a model's performance. Traditional approaches, like Shapley value, are computationally intractable for large datasets and models because they require retraining models multiple times on various data subsets. Influence-function-based methods also face significant computational hurdles. Furthermore, existing methods are often unsuitable for "partial data valuation," where one might only have access to a pre-trained model and the specific data they want to evaluate, not the entire training dataset. This gap limits practical applications in real-world data markets.
Methods
The core innovation is "Unlearning Shapley," which replaces the expensive model retraining step in traditional Shapley value calculation with machine unlearning.
- Machine Unlearning: Instead of retraining a model from scratch for each data coalition, Unlearning Shapley uses a pre-trained model and "unlearns" specific data subsets. This process aims to remove the knowledge associated with the unlearned data while preserving the model's overall capabilities.
- Approximate Unlearning: The paper employs an approximate unlearning strategy, which is more efficient than exact unlearning. This involves gradient ascent to unlearn data and two regularizers: one for output consistency (using KL divergence on logits) and another for parameter stability (using L2 distance).
- Utility Function Adaptation: The standard utility function, which measures model performance after retraining on a subset, is adapted to use the performance drop after unlearning a subset from a pre-trained model.
- Monte Carlo Sampling: For full data valuation, Monte Carlo sampling is used to estimate Shapley values efficiently by sampling data permutations.
- Partial Data Valuation: A key contribution is the ability to perform partial data valuation. This is achieved by leveraging the pre-trained model to unlearn the target dataset and comparing its performance to the original pre-trained model.
Datasets
The paper uses a combination of benchmark datasets and a large-scale text corpus for evaluation:
- Benchmark Datasets:
- FMNIST (Fashion-MNIST) for image classification.
- CIFAR-10 for image classification.
- Text Corpus: A custom corpus created by merging data from Wikitext, Fineweb, and Cosmopedia, totaling 3,000,000 text data points.
- Test Set for Text Corpus: TruthfulQA is used as the test set for text corpus data valuation.
No new benchmarks were introduced, but the authors collected and merged existing text corpora to create a large-scale dataset suitable for evaluating LLM data valuation.
Key Insights and Ideas
- Unlearning as a Shortcut: The fundamental idea that machine unlearning can serve as an efficient proxy for model retraining in data valuation is a significant insight. It drastically reduces computational overhead.
- Partial Data Valuation is Feasible: The paper demonstrates that it's possible to estimate data values even when only a pre-trained model and the target data are available, a scenario common in real-world applications.
- Trade-offs in Unlearning: While Unlearning Shapley is efficient, it's not perfect. The paper acknowledges that approximate unlearning can introduce biases, and it's less effective at detecting noisy data compared to retraining-based methods.
- Subset-level Valuation is Practical: The approach extends naturally to valuing data subsets, which is more aligned with how data is traded in markets.
Key Results
- Efficiency: Unlearning Shapley significantly reduces computational costs. For GPT2-XL on a large text corpus subset, unlearning is up to 195 times faster than training.
- Accuracy: Unlearning Shapley shows comparable performance to state-of-the-art methods (Data Shapley, Beta Shapley, Influence Function) on benchmark datasets for both point- and subset-level data valuation, particularly in data removal tasks.
- Partial Data Valuation Reliability: In simulated partial data valuation scenarios, Unlearning Shapley values show a positive correlation with exact Shapley values and a negative correlation with the performance of models trained without the target data, indicating its reliability.
- Noisy Data Detection Limitation: Unlearning Shapley is found to be less effective than retraining-based methods for noisy data detection, as unlearning incorrect knowledge might have a limited impact on performance compared to retraining.
Limitations and Future Work
- Noisy Data Detection: The paper acknowledges that Unlearning Shapley is less effective at detecting noisy data due to the nature of approximate unlearning. Future work could explore methods to improve this aspect.
- Bias in Approximate Unlearning: The authors admit that biases can arise from randomness in learning and approximate unlearning. They investigate the acceptability of this error in an appendix.
- LLM Specific Challenges: For large language models (LLMs) with huge vocabularies, unlearning can still result in high Logit Kullback-Leibler Divergence (LKD) and Knowledge Retention (KR) due to overlapping information between unlearned and remaining data. Future work could investigate techniques to mitigate this for LLMs.
- Privacy-Compliant Solutions: The paper highlights the potential of Unlearning Shapley as a privacy-compliant solution for data markets, suggesting further exploration in this direction.
Summary updated