CNN Fixations: An Unraveling Approach to Visualize the Discriminative Image Regions.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) are powerful tools in computer vision but often function as black boxes.
• Lack of transparency in CNNs hinders understanding of their decision-making processes for tasks like image classification and caption generation.

Purpose of the Study:

• To develop a method for providing visual explanations of CNN predictions.
• To enhance the interpretability of deep learning models in computer vision.

Main Methods:

• Unraveling the forward pass operation of CNNs to identify key image regions.
• Exploiting feature dependencies across network layers to pinpoint discriminative locations.
• Introducing "CNN fixations" as visual cues analogous to human eye fixations.

Main Results:

• The proposed method successfully visualizes discriminative image locations guiding CNN predictions.
• CNN fixations are generated without requiring architectural modifications, additional training, or gradient computations.
• The approach is validated across diverse CNN architectures, computer vision tasks, and data modalities.

Conclusions:

• The developed technique offers a generic and effective way to explain CNN behavior.
• This method increases the transparency of "black box" deep learning models.
• CNN fixations provide valuable insights into how models interpret visual data.