Додати до Chrome
Age has different contributions because itsglobal importance varies.
Age has different contributions because its
global importance varies.
Age has different contributions because other feature values vary.
Age has different effects because the targetdistribution varies.
Age has different effects because the target
distribution varies.
Age has different effects because the model'sbaseline varies.
Age has different effects because the model's
baseline varies.
The feature has weak individual effects but
strong global importance.
Thefeature has negligible influence on predictions across all observations.
The
feature has negligible influence on predictions across all observations.
The feature has
varying individual effects that cancel out on average.
Statements 1 and 2 only
Statements 2 and 4 only
Statements 1 and 3only
Statements 1 and 3
only
Statements 1 and 4 only
To eliminate regions with low gradient
magnitudes.
To normalize the contribution of each feature
map.
To retain regions that negatively influence the
target class.
To retain regions
that positively support the target class.
Statements 1, 2, and 3.
Statements 2 and 4.
Statements 1 and 2.
Statements 1 and 4.
The signal issuppressed because the backward signal is negative.
The signal is
suppressed because the backward signal is negative.
The signal is retained because the forwardactivation was positive.
The signal is retained because the forward
activation was positive.
The signal is suppressed because the forwardactivation was positive.
The signal is suppressed because the forward
The signal is retained because the forwardneuron was active.
neuron was active.
The score oscillates randomly around its
original value.
The score stays almost unchanged, because
important features are redundant.
The score rises steadily, sincenoise is being removed.
The score rises steadily, since
noise is being removed.
The score drops quickly,
giving a small area under the deletion curve.
I and II only
XRAI averages gradients across image segments,while IG averages gradients across baseline steps.
XRAI averages gradients across image segments,
while IG averages gradients across baseline steps.
XRAI groups pixels using their attributionmagnitudes, while IG evaluates their independent contributions.
XRAI groups pixels using their attribution
magnitudes, while IG evaluates their independent contributions.
XRAI ranks segmentedregions using IG attributions, while IG assigns attributions to individualfeatures.
XRAI ranks segmented
regions using IG attributions, while IG assigns attributions to individual
features.
XRAI computes regional gradients along theintegration path, while IG computes pixel-level gradients.
XRAI computes regional gradients along the
integration path, while IG computes pixel-level gradients.
It assigns weights based on class probability toinfluence the local surrogate.
It assigns weights based on class probability to
influence the local surrogate.
It assigns weights based on predictionconfidence to improve local accuracy.
It assigns weights based on prediction
confidence to improve local accuracy.
It assigns weights based on feature importanceto improve local fidelity.
It assigns weights based on feature importance
to improve local fidelity.
It assigns weightsbased on sample distance to influence the local surrogate.
It assigns weights
based on sample distance to influence the local surrogate.
