This repository contains the code to implement simulations in "A Network of Biologically Inspired Rectified Spectral Units (ReSUs) Learns Hierarchical Features Without Error Backpropagation".

To install requirements, run the following command:

conda create -n resu python=3.12
conda activate resu
pip install -r requirements.txt

The natural scene dataset is from Meyer et al 2014. The preprocessing of the raw images is decribed in https://elifesciences.org/articles/47579. To replicate the results in this paper

• Download the MATLAB code from https://github.com/ClarkLabCode/SynapticModel. Before running the code, the rootDataPath and sceneSourcePath directories must be set in the SetConfiguration function, or provided as arguments to that function. Then run the command config = SetConfiguration(rootDataPath, sceneSourcePath) to set the paths. Next, run params = SetModelParameters to get the parameters for the model.

• Download the Meyer 2014 dataset, which is a set of .rar archives, each containing a set of .mat files. Extract the .mat files from the archives and copy the resulting .mat files into a folder named imageData within the sceneSourcePath root directory.

• To convert the images into contrast profiles. First run ConvertNaturalScenesToContrast(config,params) function inside utlil directory. This will save the constrast as a .mat file within the root directory. Rename this .mat file as contrast_scene.mat and copy it to the data folder in this repository.

• OU process: to simulate an examplar OU process, run the notebok OU_process.ipynb. The output will be saved in the results folder. Which is used to plot Figure 2B.

• Gaussian processes: to simulate Gaussian processes with different kernels, run the notebok Gaussian_processes.ipynb. The output will be saved in the results folder. The section Switch of noise level will generate Figure S2 in the supplementary material.

• natural scene: to simulate the filters derived from CCA of contrast profiles from natural scenes, run the notebok filter_natural_scene.ipynb. The output will be saved in the results folder.

• ReSU: to simulate a two-layer ReSU network on natural scenes, run the notebook ReSU_motion_detection.ipynb. The output will be saved in the results folder.

All the figures in the main text can be reproduced by running the notebok figures_ReSU.ipynb. Note that, you have to run the previous notebooks to generate the data first.

ReSU_SM.pdf is the supplemental material for the main text.