Struct SaliencyMap 

pub struct SaliencyMap { /* private fields */ }

Saliency Maps for neural network visualization.

Computes gradients of the output with respect to input features, showing which parts of the input most affect the prediction.

Method

Uses numerical gradient estimation via finite differences:

saliency_i = ∂f/∂x_i ≈ (f(x + ε*e_i) - f(x - ε*e_i)) / (2ε)

Reference

• Simonyan, K., et al. (2014). Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps.

Implementations

impl SaliencyMap

pub fn new() -> Self

Create a new saliency map calculator.

pub fn with_epsilon(epsilon: f32) -> Self

Create with custom epsilon.

pub fn epsilon(&self) -> f32

Get epsilon value.

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

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

Compute saliency (gradient) for each input feature.

Arguments

• model_fn - Prediction function
• sample - Input to explain

Returns

Gradient values for each input feature.

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

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

Compute absolute saliency (magnitude of gradient).

Useful for visualizing which features matter, regardless of direction.

pub fn smooth_grad<F>( &self, model_fn: F, sample: &Vector<f32>, n_samples: usize, noise_level: f32, ) -> Vector<f32>

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

Compute smooth gradient by averaging over noisy samples.

SmoothGrad reduces noise in saliency maps by averaging gradients over multiple noisy versions of the input.

Reference

• Smilkov, D., et al. (2017). SmoothGrad: removing noise by adding noise.

Trait Implementations

impl Clone for SaliencyMap

fn clone(&self) -> SaliencyMap

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for SaliencyMap

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

Formats the value using the given formatter.

impl Default for SaliencyMap

fn default() -> Self

Returns the “default value” for a type.