Retina optimized for prediction across animal species

This repository contains the code of the spiking retinal model and the code for reproducing all the results reported in our preprint.

Installing dependencies

Install all required dependencies and activate the retina environment using conda.

conda env create -f environment.yml
conda activate retina

Model training

All the models can be trained by running the following script:

python scripts/model/fits.py

Alternatively, model training can be customized by directly running the training script with custom command-line arguments:

python scripts/model/train.py

Example command-line arguments modify the model architecture:

• --n_in the number of spiking units.
• --rf_size the receptive field spatial size in pixels.
• --photo_noise the added level of noise to the input clips.

or the training process:

• --n_epochs the number of training epochs.
• --batch_size the batch size.
• --lr the learning rate.

See the training script for a full list of command-line arguments.

Downloading pre-trained models

Instead of training the models from scratch, pre-trained models can be found under the releases of this repository.

Natural movie training data

We recorded various natural movies for model training. The dataset can be downloaded from: https://figshare.com/articles/dataset/Natural_movies/24265498. Make sure to set the dataset_path variable in the scripts/model/fits.py script to point to the dataset.

Reproducing paper results

All the figures in the paper can be reproduced using the notebooks found in the notebooks directory.

Reconstructing images from spike rates and latencies

To build the noisy datasets run the scripts/recon/build_noise_dataset.py script. Finally, launch the scripts/recon/fit_noise_dataset.py script to fit the readouts for reconstructing the input images using the rate and latency spike codes.

Downloading retina response datasets

All instructions for downloading and processing the retinal recordings can be found in the data/neural directory.

Predicting retinal responses from the spiking models

The script to launch each cross-validated retinal dataset fit can be found in the scripts/fits/build_and_fit directory. Scripts for determining the best performing spatial scale for each dataset are found in the scripts/fits/search directory.