Path-Sampled Integrated Gradients

2026-04-15 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Researchers have introduced Path-Sampled Integrated Gradients (PS-IG), a novel framework designed to generalize feature attribution by calculating the expected value over baselines sampled along a linear interpolation path. PS-IG is mathematically equivalent to path-weighted integrated gradients when the weighting function aligns with the sampling density's cumulative distribution function. This equivalence enables the stochastic expectation to be evaluated using a deterministic Riemann sum, which improves the error convergence rate from O(m^-1/2) to O(m^-1) for smooth models. Additionally, PS-IG analytically functions as a variance-reducing filter, strictly lowering attribution variance by a factor of 1/3 under uniform sampling, while maintaining axiomatic properties like linearity and implementation invariance.

Key takeaway

For research scientists developing or applying explainable AI methods, PS-IG offers a significant advancement in feature attribution. You should consider integrating PS-IG into your workflow to achieve more accurate and less noisy gradient-based explanations, particularly for smooth models where it promises faster error convergence and reduced variance.

Key insights

PS-IG generalizes feature attribution, improving error convergence and reducing gradient noise.

Principles

• Stochastic expectation can be evaluated deterministically.
• Variance reduction is possible while preserving axioms.

Method

PS-IG computes expected values over baselines sampled along a linear interpolation path, using a deterministic Riemann sum for evaluation.

In practice

• Apply PS-IG for improved attribution accuracy.
• Use PS-IG to reduce gradient noise in models.

Topics

• Path-Sampleed Integrated Gradients
• Feature Attribution
• Gradient Noise Reduction
• Error Convergence Rate
• Path-Weighted Integrated Gradients

Best for: Research Scientist, AI Scientist

Related on AIssential

• Path-Sampled Integrated Gradients — PS-IG unifies stochastic baseline sampling with deterministic path weighting for robust, efficient feature attribution.
• Application of integrated gradients explainability to sociopsychological semantic markers — Integrated gradients reveal word-level contributions to sociopsychological text classification, even with limited data.
• Forecasting Is Not Attribution: Localizing Decoder Bypass in Graph-Based Neural Marketing Mix Models — Forecasting accuracy in neural MMMs does not guarantee correct attribution, often hiding a "decoder bypass" failure.
• Aggregate Models, Not Explanations: Improving Feature Importance Estimation — Ensembling models directly improves feature importance estimation by reducing excess risk, especially for complex ML models.
• Explainable AI: Context-Aware Layer-Wise Integrated Gradients for Explaining Transformer Models — CA-LIG provides context-aware, layer-wise explanations for Transformer models by integrating token attributions with attention gradients.
• Local graph estimation with pathwise false discovery control — Local graph estimation with pathwise feature selection (PFS) infers interpretable substructures around target variables.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.