Natural Image Synthesis for the Retina with Variational Information Bottleneck Representation

Code for "Natural image synthesis for the retina with variational information bottleneck representation".

Description

Software Requirements

1. Tensorflow v2.3.0
2. Tensorflow-probability 0.11
3. Numpy, imageio, matplotlib, shutil, sklearn

Hardware Requirements

Our custom scripts have been executed on an NVIDIA RTX 3090 GPU.

How to Run the Code

1. Change the directory to the path of the codes where main.py exists.
2. To run the code on a trained model, make sure the models folder is downloaded and placed in the same directory as main.py. The data folder should be one level higher than main.py directory.

Enter the following commands in the command prompt:

python main.py --task test_forward # to see the spiking predictions of the IB-GP model on the Natural dataset  
python main.py --task traversal # to run the traversal on the IB-GP model  
python main.py --task train_forward # to train the IB-GP/ IB-Disjoint model  
python main.py --task adaptive_train # to train the image synthesizer  
python main.py --task latent_analyze # to visualize the autocorrelation of the latents and neural dynamics for IB-GP model

The parameters of the network, such as the beta value, the number of training epochs, and the latent dimension, can be changed by:

python main.py --task train_forward --num_epochs 1 --beta 10 --latent_dims 15

The model type can be set by:

--model_type IB-GP (default)  or IB-Disjoint

Note: The beta value should be inverted. For example, for a beta value of 0.05, set the beta to 20.

Reference

If you use this code, please cite the following paper:

@article{rahmani2022natural,
title={Natural image synthesis for the retina with variational information bottleneck representation},
author={Rahmani, Babak and Psaltis, Demetri and Moser, Christophe},
journal={Advances in Neural Information Processing Systems},
volume={35},
pages={6034--6046},
year={2022}
}