Active Continual Learning with Binarized Bayesian Neural Networks

March 27, 2026 · View on GitHub

This repository contains the code used for the experiments in Active Continual Learning with Binarized Bayesian Neural Networks. It provides a unified framework to train, evaluate, and analyze Bayesian binarized models under continual learning settings, with a strong focus on uncertainty estimation and out-of-distribution (OOD) detection.

Main contributors:

Research director:

Table of contents

Environment Setup

The repository uses a Conda environment (environment.yml). Key dependencies:

  • Conda channels: nvidia, defaults

  • Conda packages:

    • python=3.12
    • matplotlib
    • pip

  • Pip packages (installed via pip: section in environment.yml):

You can recreate the environment with:

conda env create -f environment.yml
conda activate binarized

Project Structure

The project is organized around a single main file, a set of configurations, and a collection of scripts used to reproduce all figures and appendix results.

active-continual-learning-bayesianbinn/
│
├── configurations/          # JSON configuration files (models, datasets, optimizers)
├── customLayers/            # Custom neural network layers, activations
├── datasets/                # Dataset loaders and utilities
├── figures-*/               # Output figures (main paper & appendix)
├── models/                  # Model definitions
├── optimizers/              # Optimizers and regularization methods (EWC, SI, etc.)
├── results-*/               # Serialized experiment results
├── scripts/                 # Bash scripts to reproduce figures and tables
├── utils/                   # Utility functions
│
├── main.py                  # Main training & evaluation entry point
├── environment.yml          # Conda environment specification
└── README.md                # This file

Main Training File (main.py)

The core of the project is the main.py file.

  • Loading configurations
  • Initializing datasets, models, and optimizers
  • Running the continual learning training loop
  • Exporting accuracies and uncertainty metrics

Command-line Arguments

-c, --config
    Configuration file name (without .json)

-it, --n_iterations
    Number of times to run the configuration

-v, --verbose
    Display progress bar and intermediate metrics

-ood, --ood
    Dataset for OOD detection: {fashion, pmnist, None}

-gpu, --gpu
    GPU ID to use

-wh, --weight_histogram
    Save weight histograms during training

-eln, --extract_layer_norm
    Extract layer normalization outputs

-fits, --fits_in_memory
    Whether the dataset fits in memory

-train, --train_accuracy
    Display training accuracy (requires --verbose)

-euf, --extract_uncertainties_full
    Extract epistemic uncertainty histograms per epoch

-pca, --per_class_acc
    Compute per-class accuracy

Example Usage

python main.py \
    --config main-pmnist-1000tasks-100neurons/bimu \
    --n_iterations 5 \
    --ood fashion \
    --gpu 0 \
    --verbose

Reproducing Figures and Tables

All figures and tables from the paper and appendix can be reproduced using the scripts in the scripts/ folder. Then, use the corresponding Jupyter notebooks noted appendix or main to generate the plots and tables.

Main Paper Results

  • Permuted MNIST dataset

    bash scripts/main-pmnist-table.sh

  • Animals dataset

    bash scripts/main-animals-al.sh

  • OpenLORIS dataset

    bash scripts/main-openloris-al.sh
bash scripts/main-openloris-table.sh

Appendix Experiments

  • Permuted MNIST – Memory window N

    bash scripts/appendix-pmnist-table-N.sh

  • Permuted MNIST – Activation functions

    bash scripts/appendix-pmnist-table-activation.sh

  • Permuted MNIST – Model size

    bash scripts/appendix-pmnist-table-size.sh

  • OpenLORIS – Standardized evaluation

    bash scripts/appendix-openloris-standardized-table.sh

  • OpenLORIS – Variation ratio samples

    bash scripts/appendix-openloris-al-variation-ratio.sh

  • OpenLORIS – Variation ratio dynamic threshold

    bash scripts/appendix-openloris-al-variation-threshold.sh

Each script launches a sequence of main.py runs with the appropriate configuration files and automatically stores the results.

Notes

  • Experiments can be computationally expensive; GPU usage is recommended.
  • Interrupting training safely cleans up partial results.

Citation

Please reference this work as

License

This project is licensed under the CC-BY 4.0 License - see the LICENSE file for details.