Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ADVISE: ADaptive Feature Relevance and VISual Explanations for Convolutional Neural Networks

Mohammad Mahdi Dehshibi, Mona Ashtari-Majlan, Gereziher Adhane, David Masip

TL;DR

ADVISE tackles CNN explainability by quantifying per-unit feature-map relevance through adaptive bandwidth kernel density estimation of gradients, producing class-specific saliency maps without retraining. By deriving a unit relevance score and aggregating units with the same score, it yields concise, interpretable visual explanations that highlight the most influential activation-map components. Across pretrained AlexNet, VGG16, ResNet50, and Xception on ImageNet, ADVISE generally outperforms Grad-CAM, Grad-CAM++, Score-CAM, and Layer-CAM in visual explainability while maintaining competitive time complexity, and it passes sanity checks and sensitivity axioms. The work provides a reproducible, architecture-agnostic explainability framework with a rigorous evaluation protocol, enabling more trustworthy CNN deployments and informing model design and transfer learning decisions.

Abstract

To equip Convolutional Neural Networks (CNNs) with explainability, it is essential to interpret how opaque models take specific decisions, understand what causes the errors, improve the architecture design, and identify unethical biases in the classifiers. This paper introduces ADVISE, a new explainability method that quantifies and leverages the relevance of each unit of the feature map to provide better visual explanations. To this end, we propose using adaptive bandwidth kernel density estimation to assign a relevance score to each unit of the feature map with respect to the predicted class. We also propose an evaluation protocol to quantitatively assess the visual explainability of CNN models. We extensively evaluate our idea in the image classification task using AlexNet, VGG16, ResNet50, and Xception pretrained on ImageNet. We compare ADVISE with the state-of-the-art visual explainable methods and show that the proposed method outperforms competing approaches in quantifying feature-relevance and visual explainability while maintaining competitive time complexity. Our experiments further show that ADVISE fulfils the sensitivity and implementation independence axioms while passing the sanity checks. The implementation is accessible for reproducibility purposes on https://github.com/dehshibi/ADVISE.

ADVISE: ADaptive Feature Relevance and VISual Explanations for Convolutional Neural Networks

TL;DR

ADVISE tackles CNN explainability by quantifying per-unit feature-map relevance through adaptive bandwidth kernel density estimation of gradients, producing class-specific saliency maps without retraining. By deriving a unit relevance score and aggregating units with the same score, it yields concise, interpretable visual explanations that highlight the most influential activation-map components. Across pretrained AlexNet, VGG16, ResNet50, and Xception on ImageNet, ADVISE generally outperforms Grad-CAM, Grad-CAM++, Score-CAM, and Layer-CAM in visual explainability while maintaining competitive time complexity, and it passes sanity checks and sensitivity axioms. The work provides a reproducible, architecture-agnostic explainability framework with a rigorous evaluation protocol, enabling more trustworthy CNN deployments and informing model design and transfer learning decisions.

Abstract

To equip Convolutional Neural Networks (CNNs) with explainability, it is essential to interpret how opaque models take specific decisions, understand what causes the errors, improve the architecture design, and identify unethical biases in the classifiers. This paper introduces ADVISE, a new explainability method that quantifies and leverages the relevance of each unit of the feature map to provide better visual explanations. To this end, we propose using adaptive bandwidth kernel density estimation to assign a relevance score to each unit of the feature map with respect to the predicted class. We also propose an evaluation protocol to quantitatively assess the visual explainability of CNN models. We extensively evaluate our idea in the image classification task using AlexNet, VGG16, ResNet50, and Xception pretrained on ImageNet. We compare ADVISE with the state-of-the-art visual explainable methods and show that the proposed method outperforms competing approaches in quantifying feature-relevance and visual explainability while maintaining competitive time complexity. Our experiments further show that ADVISE fulfils the sensitivity and implementation independence axioms while passing the sanity checks. The implementation is accessible for reproducibility purposes on https://github.com/dehshibi/ADVISE.
Paper Structure (8 sections, 15 equations, 5 figures, 1 table, 1 algorithm)

This paper contains 8 sections, 15 equations, 5 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Literature Review
  3. ADVISE: ADaptive VISual Explanation
  4. Experiments
  5. Evaluation Metrics
  6. Experimental result
  7. Ablation study
  8. Conclusion

Figures (5)

  • Figure 1: The visual and feature-relevance explanations for VGG16 simonyan2015very (the first row) and Xception chollet2017xception (the second row) pretrained models on ILSVRC russakovsky2015imagenet. (a, g) The input image. (b, h) The output of the Log-scaled softmax Logits for the top-5 predicted classes. (c, i) Local explanations for the prediction of the input image based on LIME ribeiro2016should. (d, j) Cumulative gradients of the last convolutional layer, where the feature map is scaled up to the resolution of the input image using bilinear interpolation. (e, k) Estimated density of the $k^{\text{th}}$ unit in feature map, which represents 2 peaks. (f, l) Adaptive cumulative gradients of units with 2 peaks in their estimated density.
  • Figure 2: (a) The $265^{\text{th}}$ unit of the activation map in the last convolution layer of the VGG16 model for the input image in Figure \ref{['fig:01a']}, where gradient values are mapped to colours in the 'cool' colour map for better visualisation. (b) Estimated kernel density with variable bandwidth (solid red line) using Eq. \ref{['eq:08']}. The grey area represents the underlying distribution of gradient values in the $265^{\text{th}}$ unit of the activation map.
  • Figure 3: The outputs of ADVISE and Grad-CAM selvaraju2017grad are compared for four images fed into the pretrained AlexNet krizhevsky2012imagenet, VGG16 simonyan2015very, ResNet50 he2016deep, and Xception chollet2017xception models on ILSVRC russakovsky2015imagenet. The use of $\phi_k(A)$ on the estimated kernel density and ADVISE show that in the explainability of (a) AlexNet prediction ('Bernese mountain dog'), two units with two peaks work better than Grad-CAM that requires 1000 units, (b) VGG16 prediction ('monastery'), four units with six peaks contribute more than Grad-CAM that requires 512 units, (c) ResNet50 prediction ('Zebra'), 177 units with one peak outperform Grad-CAM, which requires 2048 units, and (d) Xception prediction ('band aid'), eight units with three peaks perform better than Grad-CAM which utilises 2048 units.
  • Figure 4: ADVISE outputs for shallow, middle, and deep layers of (a) VGG16, (b) ResNet50, and (c) Xception pretrained models on ILSVRC.
  • Figure 5: (a) An ablated image by randomly replacing pixels with the salt and pepper noise with the noise density of $\delta =$ [0.025, 0.075, 0.125, 0.175, 0.225]. (b-e) Changes in the performance of the ADVISE and five additional visual explanation methods in AlexNet, VGG16, ResNet50, and Xception pretrained models on ILSVRC as a function of (AVX, $\delta$).