Explaining Bayesian Neural Networks (MC Dropout + UAI)

This repository contains a minimal, self-contained demo that reproduces the core idea of "Explaining Bayesian Neural Networks" on a single ImageNet example (a tiger cat image).

We take a pretrained VGG16 network, interpret its dropout layers as an approximate Bayesian posterior (MC Dropout), sample multiple explanations with LRP-CMP, and aggregate them into:

• Mean explanation: average relevance over MC samples
• Intersection (low percentile): features that are consistently relevant
• Union (high percentile): features that are relevant in at least some samples
• UAI⁺: pixelwise stability of positive relevance across samples

Everything is designed to run locally, without any Colab specific code.

---

1. Setup

1.1. Clone the repo

git clone https://github.com/your-user/explaining-bnns-demo.git
cd explaining-bnns-demo

1.2. Create and activate a virtual environment (optional but recommended)

python -m venv .venv
source .venv/bin/activate  # on Windows: .venv\Scripts\activate

1.3. Install dependencies

pip install -r requirements.txt

Note: You need a PyTorch build that matches your system. The simplest way is usually to install from pytorch.org and then add the remaining packages from requirements.txt.

---

2. Add the tiger cat image

Create the assets folder (if it does not exist yet) and place an ImageNet style tiger cat image there:

mkdir -p assets
cp /path/to/your/tiger_cat.jpg assets/tiger_cat.jpg

The notebook assumes the file exists at:

assets/tiger_cat.jpg

---

3. Running the notebook

Start Jupyter from the repository root:

jupyter lab
# or
jupyter notebook

Then open:

notebooks/01_bnn_uai_tiger_cat.ipynb

The notebook is structured into the following steps:

1. Load a pretrained VGG16 model and send it to CPU or GPU.
2. Load and preprocess assets/tiger_cat.jpg.
3. Compute a baseline explanation with LRP-CMP (no dropout at inference time).
4. Enable MC Dropout and sample multiple explanations for the same image.
5. Aggregate these explanations into:
   • Mean map
   • Intersection map (5th percentile)
   • Union map (95th percentile)
   • UAI⁺ stability map
6. Visualize all maps side by side.

---

4. Code layout

• src/models.py
  Utilities for loading the pretrained VGG16 model.

• src/explanations.py
  Implementation of LRP-CMP, normalization helpers, and an UnNormalize transform.

• notebooks/01_bnn_uai_tiger_cat.ipynb
  The main demo notebook, with explanatory text and a simple, linear workflow.

---

5. Citation

If you use this code in academic work, please consider citing:

Bykov, K. et al., "Explaining Bayesian Neural Networks" (TMLR, 2025).

This repository is a small convenience wrapper around the core ideas from that paper, focused on a single, easy to run example.