Explanation Methods List

SignXAI provides a comprehensive list of explanation methods for both PyTorch and TensorFlow models. This page details all available methods, their parameters, and framework compatibility.

Library Attribution

SignXAI builds upon two powerful explainability libraries, each providing the backend implementation for various explanation methods:

• iNNvestigate: Powers the TensorFlow implementation of Layer-wise Relevance Propagation (LRP) and other methods

  • Developed at TU Berlin by M. Alber et al.

  • Original paper: iNNvestigate neural networks!

  • Repository: github.com/albermax/innvestigate

• Zennit: Powers the PyTorch implementation of Layer-wise Relevance Propagation (LRP) variants

  • Developed at TU Berlin by C. Anders et al.

  • Original paper: Software for Dataset-wide XAI

  • Repository: github.com/chr5tphr/zennit

The SIGN method, which is SignXAI’s novel contribution, builds upon these libraries to provide improved explanations by reducing bias.

Method Overview

The table below shows all available explanation methods in SignXAI. Methods are implemented as follows:

• Gradient-based methods: Direct implementation in both PyTorch and TensorFlow

• Guided Backpropagation: Direct implementation in both PyTorch and TensorFlow

• Grad-CAM: Direct implementation in both PyTorch and TensorFlow

• LRP methods (TensorFlow): Implemented using iNNvestigate

• LRP methods (PyTorch): Implemented using Zennit

• SIGN method: Original SignXAI contribution, extending both iNNvestigate and Zennit

Available Explanation Methods

Method Name	Description	TensorFlow	PyTorch
gradient	Vanilla gradient	✓	✓
input_t_gradient	Gradient multiplied by input	✓	✓
gradient_x_sign	Gradient multiplied by sign of input	✓	✓
gradient_x_sign_mu	Gradient multiplied by sign with threshold mu	✓	✓
guided_backprop	Guided backpropagation	✓	✓
guided_backprop_x_sign	Guided backpropagation multiplied by sign	✓	✓
guided_backprop_x_sign_mu	Guided backpropagation multiplied by sign with threshold mu	✓	✓
integrated_gradients	Integrated gradients	✓	✓
smoothgrad	SmoothGrad	✓	✓
smoothgrad_x_sign	SmoothGrad multiplied by sign	✓	✓
vargrad	VarGrad (variance of gradients across noisy samples)	✓	✓
deconvnet	DeconvNet	✓	✓
grad_cam	Grad-CAM	✓	✓
grad_cam_timeseries	Grad-CAM for time series data	✓	✓
lrp_z	LRP-Z rule	✓	✓
lrpsign_z	LRP-Z with SIGN input layer rule	✓	✓
lrp_epsilon_{value}	LRP-epsilon with specified epsilon value	✓	✓
lrpsign_epsilon_{value}	LRP-epsilon with SIGN input layer rule	✓	✓
lrp_epsilon_{value}_std_x	LRP-epsilon with epsilon proportional to input std	✓	✓
lrp_alpha_1_beta_0	LRP-AlphaBeta with alpha=1, beta=0	✓	✓
lrpsign_alpha_1_beta_0	LRP-AlphaBeta with SIGN input layer rule	✓	✓
lrp_sequential_composite_a	LRP composite with layer-specific rules (variant A)	✓	✓
lrp_sequential_composite_b	LRP composite with layer-specific rules (variant B)	✓	✓

Gradient-Based Methods

Implemented directly in SignXAI with framework-specific optimizations

Vanilla Gradient

Method name: gradient

Computes the gradient of the target output with respect to the input, highlighting features that have the greatest effect on the output.

Parameters:

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

Gradient x Input

Method name: input_t_gradient

Element-wise multiplication of the gradient with the input to reduce noise and improve visualization.

Parameters:

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

Gradient x SIGN

Method name: gradient_x_sign

Multiplies the gradient with the sign of the input, focusing on the input’s direction rather than magnitude.

Parameters:

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

Method name: gradient_x_sign_mu

Includes a threshold parameter mu for more flexible sign thresholding.

Parameters:

• mu: Threshold parameter (default: 0)

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

Integrated Gradients

Method name: integrated_gradients

Computes gradients along a straight-line path from a baseline to the input to better attribute feature importance.

Parameters:

• steps: Number of steps for integration (default: 50)

• reference_inputs: Baseline input (default: zeros)

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

SmoothGrad

Method name: smoothgrad

Computes average gradients from multiple input samples with added noise to produce smoother, more visually interpretable heatmaps.

Parameters:

• augment_by_n: Number of noisy samples (default: 50)

• noise_scale: Scale of Gaussian noise (default: 0.2)

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

VarGrad

Method name: vargrad

Computes the variance of gradients across multiple noisy samples to identify unstable attributions.

Parameters:

• num_samples: Number of noisy samples (default: 50)

• noise_level: Level of Gaussian noise (default: 0.2)

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

Guided Backpropagation

Implemented directly in SignXAI with framework-specific optimizations

Method name: guided_backprop

Modifies the ReLU gradient to only pass positive gradients, producing sharper visualization.

Parameters:

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

Method name: guided_backprop_x_sign

Multiplies guided backpropagation with the sign of the input for enhanced visualization.

Parameters:

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

Method name: guided_backprop_x_sign_mu

Includes a threshold parameter mu for more flexible sign thresholding.

Parameters:

• mu: Threshold parameter

• neuron_selection: Target output neuron (class) for which to compute the gradient (default: argmax)

Grad-CAM

Implemented directly in SignXAI with framework-specific optimizations

Method name: grad_cam

Generates a localization map highlighting important regions by using the gradients flowing into the final convolutional layer.

Parameters:

• last_conv_layer_name: Name of the last convolutional layer (auto-detected if None)

• neuron_selection: Target output neuron (class) (default: argmax)

• resize: Whether to resize the output to match input dimensions (default: True)

Method name: grad_cam_timeseries

Specialized version of Grad-CAM for time series data.

Parameters:

• last_conv_layer_name: Name of the last convolutional layer (auto-detected if None)

• neuron_selection: Target output neuron (class) (default: argmax)

• resize: Whether to resize the output to match input dimensions (default: True)

Layer-wise Relevance Propagation (LRP)

TensorFlow implementation provided by iNNvestigate; PyTorch implementation provided by Zennit

LRP Base Methods

Method name: lrp_z

Basic LRP implementation following the z-rule.

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

• input_layer_rule: Rule for the input layer (default: None)

Method name: lrpsign_z

LRP-Z with SIGN input layer rule.

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

LRP with Epsilon

Methods: lrp_epsilon_{value} (e.g., lrp_epsilon_0_1, lrp_epsilon_1, etc.)

LRP with epsilon stabilization factor.

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

• input_layer_rule: Rule for the input layer (default: None)

Methods: lrpsign_epsilon_{value} (e.g., lrpsign_epsilon_0_1)

LRP-epsilon with SIGN input layer rule.

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

Methods: lrp_epsilon_{value}_std_x (e.g., lrp_epsilon_0_1_std_x)

LRP with epsilon proportional to the standard deviation of the input.

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

• input_layer_rule: Rule for the input layer (default: None)

LRP with Alpha-Beta

Method name: lrp_alpha_1_beta_0

LRP with separate treatment of positive and negative contributions.

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

• input_layer_rule: Rule for the input layer (default: None)

Method name: lrpsign_alpha_1_beta_0

LRP Alpha-Beta with SIGN input layer rule.

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

LRP Composite Rules

Method name: lrp_sequential_composite_a

LRP with layer-specific rules (variant A).

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

• input_layer_rule: Rule for the input layer (default: None)

Method name: lrp_sequential_composite_b

LRP with layer-specific rules (variant B).

Parameters:

• neuron_selection: Target output neuron (class) (default: argmax)

• input_layer_rule: Rule for the input layer (default: None)

Deep Taylor Decomposition

Method name: deep_taylor

Implements Deep Taylor decomposition using LRP epsilon as a proxy method.

Parameters:

• epsilon: Stabilizing factor for epsilon rule (default: 1e-6)

• neuron_selection: Target output neuron (class) (default: argmax)

Framework-Specific Parameters

Some parameters have different meanings or implementations between TensorFlow and PyTorch.

TensorFlow-Specific Parameters

• model_no_softmax: Model with softmax removed (done automatically)

• input_layer_rule: Input layer rule for LRP methods (‘Z’, ‘SIGN’, ‘Bounded’, ‘WSquare’, ‘Flat’)

PyTorch-Specific Parameters

• target_layer: Target layer for Grad-CAM (auto-detected if None)

• rule: LRP rule in Zennit implementation (‘epsilon’, ‘zplus’, ‘alphabeta’)

• rule_type: Advanced LRP rule type (‘alpha1beta0’, ‘epsilon’, ‘gamma’, etc.)

Common Parameters

• target_class: Target class index (used in PyTorch implementation)

• neuron_selection: Target neuron/class (used in TensorFlow implementation)

Both have the same meaning and can be used interchangeably depending on the framework.