Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Feature CAM: Interpretable AI in Image Classification

Frincy Clement, Ji Yang, Irene Cheng

TL;DR

The paper addresses the interpretability gap in image classification by introducing Feature CAM, a perturbation-activation fusion that augments Activation-Based Methods with edge-based feature descriptors from Holistically Nested Edge Detection. By generating three input variants and leveraging Grad-CAM saliency maps, Feature CAM yields fine-grained, class-discriminative visualizations that enhance human interpretability while preserving machine confidence across multiple CNN architectures. Empirical results show a 3–4× improvement in human interpretability over Grad-CAM-based baselines, with robust machine interpretability maintained, especially for lighter models. A key limitation is the dependence on Grad-CAM for localization, and future work aims to establish a stronger, model-agnostic localization baseline to broaden applicability.

Abstract

Deep Neural Networks have often been called the black box because of the complex, deep architecture and non-transparency presented by the inner layers. There is a lack of trust to use Artificial Intelligence in critical and high-precision fields such as security, finance, health, and manufacturing industries. A lot of focused work has been done to provide interpretable models, intending to deliver meaningful insights into the thoughts and behavior of neural networks. In our research, we compare the state-of-the-art methods in the Activation-based methods (ABM) for interpreting predictions of CNN models, specifically in the application of Image Classification. We then extend the same for eight CNN-based architectures to compare the differences in visualization and thus interpretability. We introduced a novel technique Feature CAM, which falls in the perturbation-activation combination, to create fine-grained, class-discriminative visualizations. The resulting saliency maps from our experiments proved to be 3-4 times better human interpretable than the state-of-the-art in ABM. At the same time it reserves machine interpretability, which is the average confidence scores in classification.

Feature CAM: Interpretable AI in Image Classification

TL;DR

The paper addresses the interpretability gap in image classification by introducing Feature CAM, a perturbation-activation fusion that augments Activation-Based Methods with edge-based feature descriptors from Holistically Nested Edge Detection. By generating three input variants and leveraging Grad-CAM saliency maps, Feature CAM yields fine-grained, class-discriminative visualizations that enhance human interpretability while preserving machine confidence across multiple CNN architectures. Empirical results show a 3–4× improvement in human interpretability over Grad-CAM-based baselines, with robust machine interpretability maintained, especially for lighter models. A key limitation is the dependence on Grad-CAM for localization, and future work aims to establish a stronger, model-agnostic localization baseline to broaden applicability.

Abstract

Deep Neural Networks have often been called the black box because of the complex, deep architecture and non-transparency presented by the inner layers. There is a lack of trust to use Artificial Intelligence in critical and high-precision fields such as security, finance, health, and manufacturing industries. A lot of focused work has been done to provide interpretable models, intending to deliver meaningful insights into the thoughts and behavior of neural networks. In our research, we compare the state-of-the-art methods in the Activation-based methods (ABM) for interpreting predictions of CNN models, specifically in the application of Image Classification. We then extend the same for eight CNN-based architectures to compare the differences in visualization and thus interpretability. We introduced a novel technique Feature CAM, which falls in the perturbation-activation combination, to create fine-grained, class-discriminative visualizations. The resulting saliency maps from our experiments proved to be 3-4 times better human interpretable than the state-of-the-art in ABM. At the same time it reserves machine interpretability, which is the average confidence scores in classification.
Paper Structure (11 sections, 9 equations, 5 figures, 3 tables)

This paper contains 11 sections, 9 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Backpropagation Based Methods (BBM)
  4. Perturbation Based Methods (PBM)
  5. Activation Based Methods (ABM)
  6. ABM as Baseline For Our Approach
  7. Proposed Method: Feature CAM
  8. Empirical Evaluation of Results
  9. Qualitative Analysis
  10. Quantitative Analysis
  11. Conclusion and Future Work

Figures (5)

  • Figure 1: Top 5 Visualizations using Feature CAM in comparison to that of Grad-CAM++ on NasNet Mobile (Top) and InceptionV3 (Bottom). Our methods show visual features of the input image better than Grad-CAM on both classifiers.
  • Figure 2: Different versions of perturbed inputs used for performing experiments
  • Figure 3: Stages of creation of an explanation map. L to R: Original Image, Saliency Map, Mask and Explanation Map across classifiers VGG16 and Xception on Grad-CAM++
  • Figure 4: A processing pipeline explaining the intuition behind Feature CAM. The top half gives an overview of the Grad-CAMs starting from extracting penultimate activations before Global Average Pooling and performing linear combinations to create saliency maps. The bottom half shows Feature CAM, its experiments and the top 5 visualizations generated using a lighter classifier (VGG16) and larger classifier (Inception V3)
  • Figure 5: Top five visualizations using Feature CAM on larger classifier - Inception ResNetV2 (left) and lighter classifier - MobileNet(right) in comparison to Existing Grad-CAM techniques