DecisionTrace

This project explores model interpretability using a minimal, from-scratch implementation of Integrated Decision Gradients (IDG) in PyTorch.

The goal is to understand how attribution maps behave across different input images and baselines, and whether the explanations align with regions of the image that influence the model’s prediction.

Contents

• min_idg_report.ipynb: Experimental report and qualitative analysis of attribution behavior
• minimal_idg_implementation.ipynb: Step-by-step implementation of Integrated Decision Gradients
• min_IDG_images/: Image inputs used for attribution experiments

How to View

The notebooks can be viewed directly on GitHub, or run locally with:

pip install torch torchvision matplotlib pillow
jupyter notebook