Interpreting Low-Level Vision Models With Causal Effect Maps

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks excel at low-level vision but lack interpretability.
• Understanding deep models is crucial for network design and reliability.

Purpose of the Study:

• Introduce causality theory to interpret low-level vision models.
• Propose a model- and task-agnostic method, Causal Effect Map (CEM).
• Visualize and quantify input-output relationships (positive/negative effects).

Main Methods:

• Applied Causal Effect Map (CEM) to analyze various low-level vision tasks.
• Utilized causality theory for model interpretation.

Main Results:

• Larger receptive fields do not always improve performance.
• Global receptive field mechanisms (e.g., channel attention) may be ineffective for image denoising.
• Multi-task training can lead networks to favor local over global information.

Conclusions:

• CEM provides a novel diagnostic tool for deep vision models.
• Findings challenge common assumptions about information processing in deep vision.
• The method offers deeper insights into low-level vision model behavior.