pulsatrix 1.1.0
Loading...
Searching...
No Matches
pulsatrix::ExplanationCheckOptions Struct Reference

How BuildExplanation() runs the checks. The defaults are the published ones where there is one; each is recorded in Explanation::settings. More...

#include <explanation_report.hpp>

Collaboration diagram for pulsatrix::ExplanationCheckOptions:

Public Attributes

bool required_only = false
 Run only the checks the reason requires, and mark the rest NotRun.
 
uint64_t seed = 0
 Seed for the random orderings, the input noise and the re-initialized weights.
 
int64_t faithfulness_steps = 20
 Faithfulness: points on the deletion and insertion curves.
 
Imputation imputation = Imputation::Constant(0.0f)
 Faithfulness: how removed features are filled. Use Imputation::NoisyLinear (ROAD) for images.
 
int64_t random_orderings = 19
 Faithfulness: random feature orderings to compare with. With 19, passing means beating every one of them, p = 0.05.
 
double alpha = 0.05
 Faithfulness: the largest p-value that passes.
 
int64_t stability_repeats = 2
 Stability: extra runs on the same input.
 
int64_t perturbations = 3
 Stability: runs on the input plus Gaussian noise, skipped for token and symbolic input.
 
double perturbation_scale = 0.01
 Stability: the noise's standard deviation, as a fraction of the input's.
 
double min_stability = 0.8
 Stability: the least Spearman correlation of attribution magnitudes that passes, both between repeated runs and under input noise.
 
double robustness_radius = 0.05
 Robustness: perturbations sampled uniformly within this fraction of the input's spread (L∞), kept when they keep the predicted class.
 
int64_t robustness_samples = 10
 
double max_sensitivity = 0.5
 Robustness: the largest relative change of the explanation that passes, both the largest over the samples (max-sensitivity) and at FGSM's point.
 
double conservation_tolerance = 0.05
 Conservation: the largest gap between the attributions' sum and the output they explain, as a fraction of that output.
 
double max_random_similarity = 0.5
 Randomization and null model: the largest Spearman correlation between the trained and the randomized model's attribution magnitudes that passes.
 

Detailed Description

How BuildExplanation() runs the checks. The defaults are the published ones where there is one; each is recorded in Explanation::settings.

Member Data Documentation

◆ alpha

double pulsatrix::ExplanationCheckOptions::alpha = 0.05

Faithfulness: the largest p-value that passes.

◆ conservation_tolerance

double pulsatrix::ExplanationCheckOptions::conservation_tolerance = 0.05

Conservation: the largest gap between the attributions' sum and the output they explain, as a fraction of that output.

◆ faithfulness_steps

int64_t pulsatrix::ExplanationCheckOptions::faithfulness_steps = 20

Faithfulness: points on the deletion and insertion curves.

◆ imputation

Imputation pulsatrix::ExplanationCheckOptions::imputation = Imputation::Constant(0.0f)

Faithfulness: how removed features are filled. Use Imputation::NoisyLinear (ROAD) for images.

◆ max_random_similarity

double pulsatrix::ExplanationCheckOptions::max_random_similarity = 0.5

Randomization and null model: the largest Spearman correlation between the trained and the randomized model's attribution magnitudes that passes.

◆ max_sensitivity

double pulsatrix::ExplanationCheckOptions::max_sensitivity = 0.5

Robustness: the largest relative change of the explanation that passes, both the largest over the samples (max-sensitivity) and at FGSM's point.

◆ min_stability

double pulsatrix::ExplanationCheckOptions::min_stability = 0.8

Stability: the least Spearman correlation of attribution magnitudes that passes, both between repeated runs and under input noise.

◆ perturbation_scale

double pulsatrix::ExplanationCheckOptions::perturbation_scale = 0.01

Stability: the noise's standard deviation, as a fraction of the input's.

◆ perturbations

int64_t pulsatrix::ExplanationCheckOptions::perturbations = 3

Stability: runs on the input plus Gaussian noise, skipped for token and symbolic input.

◆ random_orderings

int64_t pulsatrix::ExplanationCheckOptions::random_orderings = 19

Faithfulness: random feature orderings to compare with. With 19, passing means beating every one of them, p = 0.05.

◆ required_only

bool pulsatrix::ExplanationCheckOptions::required_only = false

Run only the checks the reason requires, and mark the rest NotRun.

◆ robustness_radius

double pulsatrix::ExplanationCheckOptions::robustness_radius = 0.05

Robustness: perturbations sampled uniformly within this fraction of the input's spread (L∞), kept when they keep the predicted class.

◆ robustness_samples

int64_t pulsatrix::ExplanationCheckOptions::robustness_samples = 10

◆ seed

uint64_t pulsatrix::ExplanationCheckOptions::seed = 0

Seed for the random orderings, the input noise and the re-initialized weights.

◆ stability_repeats

int64_t pulsatrix::ExplanationCheckOptions::stability_repeats = 2

Stability: extra runs on the same input.


The documentation for this struct was generated from the following file: