README.md

CUE Meets LRP

Exploring how different LRP composite rules explain the same prediction, evaluated through the CUE model lens (Labarta et al., 2025).

What it does

1. Loads VGG16 and classifies one image
2. Applies three zennit composite rules to produce heatmaps
3. Scores each on CUE dimensions (Legibility, Readability, Interpretability) using simple proxies
4. Asks Qwen2.5-VL (VLM) for a second perspective on each explanation

Run locally

pip install zennit torchvision transformers accelerate qwen-vl-utils
jupyter notebook cue_meets_lrp.ipynb

Or click the Colab badge above for zero-setup execution.

Vincent Lange · vincentelange@gmail.com