Robust Neural Networks

This repository contains a collection or robust neural network architectures developed at the Australian Centre For Robotics (ACFR). All networks are implemented in Python/JAX.

Implemented network architectures include:

• Lipschitz-bounded Sandwich MLPs from Wang & Manchester (ICML 2023). Tutorial:

• Recurrent Equilibrium Network (REN) from Revay, Wang, & Manchester (TAC 2023).

• [WIP] Monotone, Bi-Lipschitz (BiLipNet), and Polyak-Lojasiewicz networks (PLNet) from Wang, Dvijotham, & Manchester (ICML 2024).

• Robust Recurrent Deep Network (R2DN) from Barbara, Wang, & Manchester (arXiv 2025).

This repository is a work-in-progress. More network architectures, tutorials, and documentation will be added as we go along.

Installation for Development

To install the required dependencies, open a terminal in the root directory of this repository and enter the following commands.

    ./install.sh

This will create a Python virtual environment at ./venv and install all dependencies. If you would rather create a virtual environment with conda, poetry, or something else, feel free to modify the install.sh script.

A Note on Dependencies

All code was tested and developed in Ubuntu 22.04 with CUDA 12.4 and Python 3.10.12.

Requirements were generated with pipreqs. The install.sh will check for whether CUDA is available for your machine, and install the corresponding jax package.

Running an Example

Once you have installed the package as above, simply activate the virtual environment and run any of the scripts in the examples/ folder. For example, from the root directory of the project, run:

source venv/bin/activate
python examples/sandwich_mnist.py

Contact

Please contact Nicholas Barbara ([email protected]) with any questions.