Pre-training artificial neural networks with spontaneous retinal activity improves motion prediction in natural scenes

In this repository, you'll find the code used to generate the results presented in our publication:

May, Lilly, Alice Dauphin, and Julijana Gjorgjieva. "Pre-training artificial neural networks with spontaneous retinal activity improves motion prediction in natural scenes." PLOS Computational Biology 21.3 (2025): e1012830..

We provide scripts for generating the data, training the artificial neural networks (ANNs), and evaluating the performance and characteristics of the trained ANNs.

📂 Datasets

This repository includes code to generate the following datasets, located in the dataset folder:

• Virtual corridor simulation: A dataset of a natural scene with prominent optic flow, created using the 3D animation software Blender and its Python API. This dataset mimics the experience of a mouse navigating through a corridor with visual cues on the walls.
• Retinal waves: A dataset of spontaneous retinal activity, also known as retinal waves, which propagate across the retina with a directional bias. The dataset was generated based on the retinal wave model introduced by Teh et al. (2023).

📦 Accessing the Datasets: The pre-generated datasets are available via Zenodo.

⚙️ ANN models

We provide code to train (model/train_cli.py) and evaluate (evaluation folder) ANN models designed for the task of Next-Frame Prediction.

Our implementation is based on convolutional recurrent neural networks, with a strong focus on modularity. For example, our models support interchangeable recurrent layers, including:

• Vanilla RNN
• GRU
• LSTM

This flexible architecture makes it easy to experiment with different configurations and evaluate their performance.

💬 Got questions or ran into issues?

Open an issue in this repo or contact Lilly - happy to help!