On the Accuracy of Influence Functions for Measuring Group Effects

Abstract

Influence functions estimate the effect of removing a training point on a model without the need to retrain. They are based on a first-order Taylor approximation that is guaranteed to be accurate for sufficiently small changes to the model, and so are commonly used to study the effect of individual points in large datasets. However, we often want to study the effects of large groups of training points, e.g., to diagnose batch effects or apportion credit between different data sources. Removing such large groups can result in significant changes to the model. Are influence functions still accurate in this setting? In this paper, we find that across many different types of groups and for a range of real-world datasets, the predicted effect (using influence functions) of a group correlates surprisingly well with its actual effect, even if the absolute and relative errors are large. Our theoretical analysis shows that such strong correlation arises only under certain settings and need not hold in general, indicating that real-world datasets have particular properties that allow the influence approximation to be accurate.

References in corpus (8)

• Understanding Black-box Predictions via Influence Functions
• Understanding the Origins of Bias in Word Embeddings
• Stronger Data Poisoning Attacks Break Data Sanitization Defenses
• Interpreting Black Box Predictions using Fisher Kernels
• Contamination Attacks and Mitigation in Multi-Party Machine Learning
• A Swiss Army Infinitesimal Jackknife
• A scalable estimate of the extra-sample prediction error via approximate leave-one-out
• A Higher-Order Swiss Army Infinitesimal Jackknife

Cited by in corpus (31)

• Fast Yet Effective Machine Unlearning
• Training Data Influence Analysis and Estimation: A Survey
• Approximate Data Deletion from Machine Learning Models
• Markov Chain Monte Carlo-Based Machine Unlearning: Unlearning What Needs to be Forgotten
• HYDRA: Hypergradient Data Relevance Analysis for Interpreting Deep Neural Networks
• Not All Unlabeled Data are Equal: Learning to Weight Data in Semi-supervised Learning
• Hessian-based toolbox for reliable and interpretable machine learning in physics
• Cross-validation Confidence Intervals for Test Error
• Influence Functions in Deep Learning Are Fragile
• Causal Interpretability for Machine Learning -- Problems, Methods and Evaluation
• Rethinking Influence Functions of Neural Networks in the Over-parameterized Regime
• Understanding Instance-based Interpretability of Variational Auto-Encoders
• RelatIF: Identifying Explanatory Training Examples via Relative Influence
• On Second-Order Group Influence Functions for Black-Box Predictions
• Chasing Your Long Tails: Differentially Private Prediction in Health Care Settings
• FairIF: Boosting Fairness in Deep Learning via Influence Functions with Validation Set Sensitive Attributes
• The Thousand Faces of Explainable AI Along the Machine Learning Life Cycle: Industrial Reality and Current State of Research
• DeepSeer: Interactive RNN Explanation and Debugging via State Abstraction
• Influence Function based Data Poisoning Attacks to Top-N Recommender Systems
• Revisiting the Fragility of Influence Functions
• Understanding Goal-Oriented Active Learning via Influence Functions
• Interactive Label Cleaning with Example-based Explanations
• FastIF: Scalable Influence Functions for Efficient Model Interpretation and Debugging
• Approximate Cross-Validation for Structured Models
• Using Wavelets to Analyze Similarities in Image-Classification Datasets
• DANCE: Enhancing saliency maps using decoys
• Quantifying Epistemic Uncertainty in Deep Learning
• Differential Privacy, Linguistic Fairness, and Training Data Influence: Impossibility and Possibility Theorems for Multilingual Language Models
• ModelPred: A Framework for Predicting Trained Model from Training Data
• Model-Agnostic Explanations using Minimal Forcing Subsets
• Approximate Cross-Validation with Low-Rank Data in High Dimensions