Abstract
Influence Functions (IFs) are a cornerstone of explainable AI, designed to identify which training samples most significantly impact a model's predictions. However, their application to deep learning has been historically "fragile," often yielding unreliable results in over-parameterized regimes. This paper establishes a rigorous theoretical foundation for IFs using Neural Tangent Kernel (NTK) theory. We prove that for wide neural networks, NTK-based influence estimates achieve high fidelity to true leave-one-out influence. Furthermore, we provide a mathematical explanation for the failure of traditional estimation methods, identifying weak regularization and low data density as key sources of error.
Introduction
A fundamental question in machine learning interpretability is: "How would a model's prediction change if a specific training point were removed?" Influence Functions (IFs) attempt to answer this without the prohibitive cost of retraining the model for every data point. The standard approach, introduced by Koh and Liang (2017), utilizes the Implicit Hessian-Vector Product (IHVP). Despite its popularity, IHVP often fails in deep learning contexts. This work addresses this gap by leveraging the properties of the over-parameterized regime to provide a more robust and theoretically grounded influence estimation method.
Methodology: The NTK Perspective
The authors bridge the gap between complex neural networks and tractable mathematical models using Neural Tangent Kernel (NTK) theory. An infinitely wide neural network trained via gradient descent behaves like a linear model performing kernel regression.
- NTK-based IF Calculation: Instead of the non-linear Hessian of a standard network, influence is calculated for the network's "NTK twin."
- Theoretical Guarantees: Theorems 1 and 2 provide the first rigorous proof that as the width of a two-layer ReLU network increases, the error between NTK-based influence and true "leave-one-out" influence vanishes.
Analyzing the Fragility of Influence Functions
A major contribution of this work is the derivation of mathematical bounds to analyze why traditional IHVP methods often fail.
- Regularization as a Shield: The accuracy of IHVP estimation is lower-bounded by a term inversely related to the regularization parameter ($\lambda$). When regularization is weak, the Hessian becomes "noisy," rendering influence estimates meaningless.
- The "Density" Trap: Estimation error is also linked to data distribution; IHVP is significantly less accurate for training points located in low-density regions (outliers).
Experimental Results
The theoretical findings were validated on MNIST and CIFAR-10 datasets using networks with widths up to $8 \times 10^4$ neurons.
- High Fidelity: The NTK-based influence estimates demonstrated an exceptional correlation with actual retraining results. For the widest networks, the Pearson correlation (R) exceeded 0.99 on both benchmarks.
- Width and Convergence: Approximation error decreased consistently as hidden layer width increased, confirming the NTK-based theoretical predictions.
Key Insights
- Influential = Complex: The most influential samples are those that increase model complexity within the NTK framework. Redundant or "simple" samples tend to be uninfluential.
- Dynamic Influence: The NTK approach allows for tracking influence throughout the training process. Early training is dominated by "clean" samples, while "noisy" or mislabeled samples become more influential later, providing a theoretical basis for the effectiveness of early stopping.
Conclusion
This research provides the necessary theoretical scaffolding to use Influence Functions reliably in modern deep learning. By moving from the fragile IHVP method to an NTK-based approach, researchers can gain more accurate insights into model behavior, particularly in the over-parameterized regime. Future work will focus on extending these proofs to deeper architectures like ResNets and Transformers.
Summary updated