Integrated Gradients

Integrated Gradients is the unique path method that satisfies three essential axioms simultaneously, making it a theoretically sound choice for high-stakes explanations:

• Completeness: Attributions sum to the output difference from baseline, ensuring no importance is lost or created
• Sensitivity-n: If a feature's value changes and the model output changes, that feature receives non-zero attribution
• Implementation Invariance: Two functionally identical networks produce identical attributions regardless of internal architecture differences

These properties make Integrated Gradients admissible under the axiomatic attribution framework, unlike heuristic methods such as Guided Backpropagation which violate sensitivity.

Simple Gradient × Input methods fail when networks saturate—features that strongly activate the correct class receive near-zero gradients, making them appear falsely unimportant. Integrated Gradients overcomes this by:

• Accumulating gradients at many points along the interpolation path, not just at the final input
• Capturing the non-linear transition where features shift from baseline to signal
• Providing meaningful attributions even in flat regions of the loss landscape

This makes it especially valuable for deep networks where shattered gradients and saturation are common, producing coherent saliency maps where simple gradients produce noise.

The choice of baseline fundamentally shapes the explanation. The baseline represents the absence of information and must be carefully chosen for the domain:

• Vision models: A black image, Gaussian noise, or blurred version of the input
• NLP models: Zero embedding vectors or padding token sequences
• Tabular data: Median feature values or domain-specific neutral references

Poor baseline selection leads to attribution artifacts—features similar between baseline and input receive zero attribution even if important. Best practice involves testing multiple baselines and using Expected Gradients to average over a distribution of references.

The exact integral is intractable, so practical implementations use Riemann approximation with discrete steps. Key implementation considerations include:

• Number of steps: Typically 20-300 steps along the path; more steps improve accuracy at computational cost
• Gaussian noise: Adding small perturbations at each step can smooth attributions (related to SmoothGrad)
• Batch processing: Steps can be computed in parallel for GPU efficiency

The approximation error decreases as the number of steps increases, with 50-100 steps providing a good balance between attribution fidelity and inference time for most production deployments.

The Integrated Gradients framework has spawned several important extensions for specific use cases:

• Integrated Hessians: Captures feature interactions using second-order derivatives, revealing which features work together
• Neuron Conductance: Decomposes attribution through hidden neurons to understand internal representations
• Expected Gradients: Averages over a distribution of baselines to remove baseline selection bias and unifies Integrated Gradients with SHAP
• Blur Integrated Gradients: Uses progressively blurred versions of the input as the path, avoiding artifacts from black baselines

These variants extend the core axiomatic guarantees to richer explanation types.

The mathematical foundation that evaluates feature attribution methods based on their adherence to specific axioms. Integrated Gradients is unique in satisfying Completeness, Sensitivity-n, and Implementation Invariance simultaneously.

• Completeness: Attributions sum to the output difference from a baseline.
• Sensitivity-n: A feature with zero effect gets zero attribution.
• Implementation Invariance: Identical functions yield identical attributions regardless of architecture.

An attribution method that unifies Integrated Gradients and SHAP by averaging gradients over a distribution of background samples rather than a single baseline.

• Eliminates the need to select a single, often arbitrary, reference point.
• Computes the expected value of the gradient across the training data distribution.
• Provides a more robust and less baseline-sensitive feature importance score.

A critical failure mode of raw gradient methods that Integrated Gradients solves. When a feature strongly activates the correct class, the gradient can saturate to near-zero, making the feature appear falsely unimportant.

• Integrated Gradients accumulates gradients along the path, capturing the feature's effect before saturation occurs.
• This is why simple Gradient × Input often fails on deep networks.

A general class of attribution techniques that define importance by integrating gradients along a specified path from a baseline to the input. Integrated Gradients is the canonical straight-line path method.

• The choice of path defines the attribution distribution.
• Non-linear paths can encode specific prior beliefs about feature interactions.
• All path methods satisfy the completeness axiom by construction.

An extension of Integrated Gradients that measures the importance of a specific hidden neuron rather than an input feature.

• Computes the flow of the gradient through a target neuron along the integration path.
• Enables debugging of internal network components by identifying which neurons are most responsible for a prediction change.
• Bridges input attribution with mechanistic interpretability.

An extension that uses the second-order Hessian matrix to capture feature interactions and curvature, providing a more complete attribution for highly non-linear models.

• Standard Integrated Gradients captures only first-order effects.
• Integrated Hessians decomposes the output into pairwise feature interaction terms.
• Essential for understanding complex feature synergies in deep networks.