Human-in-the-Loop Optimization for Deep Stimulus Encoding in Visual Prostheses

This repository houses code accompanying the paper:

Granley, J., Fauvel, T., Chalk, M., & Beyeler, M. (2023). Human-in-the-Loop Optimization for Deep Stimulus Encoding in Visual Prostheses. Proceedings of the Thirty-seventh Conference on Neural Information Processing Systems (link)

The main demo notebook is code/HILO.ipynb

Instructions for Installation

TLDR Installation

Install Matlab R2023a with deep learning, computer vision, and machine learning and statistics toolboxes

pip install -r requirements.txt
pip install git+https://github.com/pulse2percept/pulse2percept@410ece
pip install matlabengine==9.14.3
cd code/matlab
git clone [email protected]:TristanFauvel/BO_toolbox.git
git clone [email protected]:TristanFauvel/GP_toolbox.git

Full instructions and troubleshooting

This code was developed using Python 3.10; it might work with previous versions, but if you run into any errors, try Python 3.10.

Included is a requirements.txt file, which can be run to obtain all of the python dependencies. Also included is a python_env.txt for reference, with a complete list of all python packages and versions installed.

pip install -r requirements.txt

This package also uses the following software which require additional installation steps:

• pulse2percept. This is an open source package for simulating visual prostheses. Any new version should work. To get the latest version, run pip install git+https://github.com/pulse2percept/pulse2percept. If you run into issue, the exact version used for this paper was commit 202fac. This likely will install without issue. But it also depends on some non-python packages (e.g. gcc). Many systems have these installed already, but if your installation fails, please go through the installation instructions for pulse2percept at https://pulse2percept.readthedocs.io/en/stable/install.html
• Tensorflow. Tensorflow can also be installed with pip (we used version 2.12.0, listed in requirements.txt), but you may have to set up CUDA or install differently if you want it to use a GPU. See https://www.tensorflow.org/install/pip
• Matlab. All Bayesian optimization is performed in Matlab, using existing open source Bayesian optimization and Gaussian Process toolboxes (https://github.com/TristanFauvel/BO_toolbox, https://github.com/TristanFauvel/GP_toolbox). If you do not have, or do not wish to install Matlab, then you can still run all code for the deep stimulus encoder and phosphene model, and see the results from HILO in HILO.ipynb from us running the notebook. The following section provides more details on installing Matlab.

Matlab installation

Matlab is not freely available, however, most universities have licensing deals to get it for free. We used matlab version R2023a, but other versions might also work. Required matlab toolboxes (provided with matlab as additional downloads):

• deep learning toolbox,
• computer vision toolbox
• statistics and machine learning toolbox

Once installed, there are two additional steps that must be performed to set up matlab python extension and install open source toolboxes:

1. Install the python matlab engine. If you have R2023a: pip install matlabengine==9.14.3. Otherwise, see directions at https://www.mathworks.com/help/matlab/matlab_external/install-the-matlab-engine-for-python.html.
2. Download the external, open source Bayesian optimization and Gaussian process toolboxes to the code/matlab folder.
   cd code/matlab
git clone [email protected]:TristanFauvel/BO_toolbox.git
git clone [email protected]:TristanFauvel/GP_toolbox.git