Struct IntegratedGradients 

pub struct IntegratedGradients { /* private fields */ }

Integrated Gradients for neural network attribution.

Computes attributions by integrating gradients along a path from a baseline to the input.

IG_i(x) = (x_i - x'_i) * ∫ (∂F(x' + α(x-x')) / ∂x_i) dα

where x’ is a baseline (typically zeros).

References

• Sundararajan, M., et al. (2017). Axiomatic Attribution for Deep Networks.

Implementations

impl IntegratedGradients

pub fn new(n_steps: usize) -> IntegratedGradients

Create new Integrated Gradients explainer.

Arguments

• n_steps - Number of steps for Riemann approximation (default: 50)

pub fn attribute<F>( &self, model_fn: F, sample: &Vector<f32>, baseline: &Vector<f32>, ) -> Vector<f32>

where F: Fn(&Vector<f32>) -> f32,

Compute attributions using numerical gradient approximation.

Arguments

• model_fn - Prediction function
• sample - Input to explain
• baseline - Baseline input (typically zeros)

Returns

Attribution for each input feature

Trait Implementations

impl Debug for IntegratedGradients

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Default for IntegratedGradients

fn default() -> IntegratedGradients

Returns the “default value” for a type.