Course contentsShow
← Machine Learning and Deep Learning
Lesson 3253 of 3,53871. Interpretability: Neural Network MethodsFree lesson

Variants: Expected Gradients and Blur IG

Extensions using multiple baselines or distributions to capture uncertainty and improve attribution quality.

Variants: Expected Gradients and Blur IG

What You'll Learn

You'll discover how modern extensions of Integrated Gradients use multiple baselines or distributions to produce more robust, uncertainty-aware attributions.

The Core Problem with Single Baselines

Standard Integrated Gradients uses one baseline (like an all-black image or zero vector). But what if your choice of baseline is arbitrary or biased? Different baselines can yield different attributions, raising the question: which one is "correct"?

Expected Gradients: Distributional Baselines

Expected Gradients replaces a single baseline with a distribution of baselines, typically sampled from your training data. Instead of integrating from one reference point, you:

  1. Sample multiple baselines from a representative distribution
  2. Compute path integrals from each baseline to your input
  3. Average the resulting attributions

This approach captures uncertainty in baseline selection and produces attributions that reflect typical model behavior across realistic reference points. Think of it like getting multiple second opinions from different starting perspectives rather than relying on one arbitrary viewpoint.

Blur IG: Uncertainty Through Image Blur

Blur Integrated Gradients takes a different angle for image models. Instead of a discrete baseline, it creates a continuum of increasingly blurred versions of the input image itself. The path goes from maximally blurred (baseline) to sharp (actual input).

This is particularly intuitive: blur removes fine details progressively, letting you see which features emerge as most critical when detail is restored. The baseline is inherently image-dependent and meaningful.

Why These Variants Matter

Both methods improve attribution quality by:

  • Reducing sensitivity to arbitrary baseline choices
  • Incorporating uncertainty and variability
  • Producing more stable, reliable explanations across different inputs

Key Takeaway: Expected Gradients and Blur IG extend Integrated Gradients by using multiple or distribution-based baselines, capturing uncertainty and producing more robust, contextually-aware feature attributions.