Training Data Attribution via Approximate Unrolled Differentiation
This work addresses the challenge of accurately attributing model behavior to training data for machine learning practitioners, though it appears incremental as it builds on existing TDA methods.
The paper tackles the problem of training data attribution (TDA) by introducing Source, an approximate unrolling-based method that combines the benefits of implicit-differentiation and unrolling approaches, resulting in improved performance in counterfactual prediction, especially in non-converged models and multi-stage pipelines.
Many training data attribution (TDA) methods aim to estimate how a model's behavior would change if one or more data points were removed from the training set. Methods based on implicit differentiation, such as influence functions, can be made computationally efficient, but fail to account for underspecification, the implicit bias of the optimization algorithm, or multi-stage training pipelines. By contrast, methods based on unrolling address these issues but face scalability challenges. In this work, we connect the implicit-differentiation-based and unrolling-based approaches and combine their benefits by introducing Source, an approximate unrolling-based TDA method that is computed using an influence-function-like formula. While being computationally efficient compared to unrolling-based approaches, Source is suitable in cases where implicit-differentiation-based approaches struggle, such as in non-converged models and multi-stage training pipelines. Empirically, Source outperforms existing TDA techniques in counterfactual prediction, especially in settings where implicit-differentiation-based approaches fall short.