Sridhar et al., 2025

Dataset at a glance

Retinal piece	Species	Recording technique	Stimulus type(s)	Stim channels	Total cells (not quality filtered)	Original recording fs	Original stim fs	Test repeats	Train repeats	Pixel size on retina [µm]
Retina 01	marmoset	multielectrode arrays (MEAs)	white noise (WN) & natural movie (NM)	1 for both WN and NM	370	85	85	WN: 11, NM: 10	WN: 11, NM: 21	WN: 30, NM: 7.5 (downsampled to 30 for training)
Retina 02	marmoset	multielectrode arrays (MEAs)	white noise (WN) & natural movie (NM)	1 for both WN and NM	181	85	85	WN: 10, NM: 11	WN: 10, NM: 11	WN: 30, NM: 7.5 (downsampled to 30 for training)
Retina 03	marmoset	multielectrode arrays (MEAs)	white noise (WN) & natural movie (NM)	1 for both WN and NM	495	85	85	WN: 10, NM: 20	WN: 10, NM: 20	WN: 30, NM: 7.5 (downsampled to 30 for training)
Retina 04	marmoset	multielectrode arrays (MEAs)	white noise (WN) & natural movie (NM)	1 for both WN and NM	110	85	85	WN: 10, NM: 20	WN: 10, NM: 20	WN: 30, NM: 7.5 (downsampled to 30 for training)

Dataset description

Spike train data from marmoset retinal ganglion cells under visual stimulation: The dataset contains spike times of different types of ganglion cells, recorded extracellularly with multielectrode arrays from four isolated retinas of marmoset monkeys.

The visual stimuli shown are spatio-temporal white noise and naturalistic stimuli. Each of these stimuli is stored as a separate dataset having its own dataset class within open-retina.

For natural movies, use the MarmosetMovieDataset class in openretina/data_io/sridhar_2025/dataloaders.py.

For white noise, use the NoiseDataset class in openretina/data_io/sridhar_2025/dataloaders.py.

Natural movie dataset description

The naturalistic movie stimulus was derived from the film Tears of Steel by Blender Studios. It was processed to grayscale and projected at a resolution of 1600x1200 pixels. For model training, it is cropped to 800x600 pixels by default and downsampled to 200x150, matching the white-noise dataset resolution. Jittering was introduced to the movie frames to mimic the statistics of active vision in marmosets, including saccades, fixations, and fixational eye movements.

Dataset structure

This dataset is organized into five main directories containing different types of data: * fixations: Contains text files with eye fixation data, including original and 4x downsampled versions for different seeds. The complete_fixations_seed_2022and2023_s4_fixed.txt file is used as the default for model training. * responses: Stores .pkl (pickle) files containing recorded cell responses to fixation movies. * stas: Holds .npy (NumPy) files for spike-triggered averages (STAs), indexed by both the retinal piece dataset index (e.g., 01, 02) and individual cell index (e.g., cell_0 to cell_370). * stimuli_padded: Contains the original padded stimulus images as .npy files. * stimuli_padded_4: Contains the 4x downsampled version of the padded stimulus images as .npy files, corresponding to the fixations in files like complete_fixations_seed_2022and2023_s4_fixed.txt.

nm_marmoset_data
├── fixations
│   ├── complete_fixations_seed_2022and2023_s4_fixed.txt (Default for complete model training, 4x downsampled)
│   ├── complete_fixations_seed_2022.txt (Original fixation file for seed 2022)
│   ├── complete_fixations_seed_2023.txt (Original fixation file for seed 2023)
│   ├── complete_fixations_seed_2022_s4.txt (Seed 2022 fixations for 4x downsampled stimulus)
│   └── complete_fixations_seed_2023_s4.txt (Seed 2023 fixations for 4x downsampled stimulus)
├── responses
│   ├── cell_responses_01_fixation_movie.pkl
│   ├── cell_responses_02_fixation_movie.pkl
│   ├── cell_responses_03_fixation_movie.pkl
│   └── cell_responses_04_fixation_movie.pkl
├── stas
│   ├── cell_data_01_WN_stas_cell_0.npy
│   ├── cell_data_01_WN_stas_cell_1.npy
│   ├── ...
│   └── cell_data_01_WN_stas_cell_370.npy
│   ├── cell_data_02_WN_stas_cell_0.npy
│   └── ...
├── stimuli_padded
│   ├── 00000_img_10088.npy
│   ├── 00001_img_10089.npy
│   ├── ...
│   └── 14543_img_14758.npy
└── stimuli_padded_4
    ├── 00000_img_10088.npy
    ├── 00001_img_10089.npy
    ├── ...
    └── 14543_img_14758.npy

White noise dataset description

The white noise stimulus consists of a sequence of frames of black and white squares (100% Michelson contrast) arranged in a uniform checkerboard pattern.

Spatial layout: 150x200 stimulus squares, with each square spanning 30 µm x 30 µm on the retinal surface.

Dataset structure

wn_marmoset_data
├── responses
│   ├── cell_responses_01_wn.pkl
│   ├── cell_responses_02_wn.pkl
│   ├── cell_responses_03_wn.pkl
│   └── cell_responses_04_wn.pkl
├── stas
│   ├── cell_data_01_WN_stas_cell_0.npy
│   ├── cell_data_01_WN_stas_cell_1.npy
│   ├── ...
│   └── cell_data_01_WN_stas_cell_370.npy
│   ├── cell_data_02_WN_stas_cell_0.npy
│   └── ...
├── non_repeating_stimuli_1 # white noise stimuli shown to Retina 01, 02 and 03
│   ├── trial_000
│       └── all_images.npy
│   ├── ...
│   └── trial_024.npy
|       └── all_images.npy
├── non_repeating_stimuli_2 # white noise stimuli shown to Retina 04
│   ├── trial_000
│       └── all_images.npy
│   ├── ...
│   └── trial_024.npy
|       └── all_images.npy
├── repeating_stimuli_1  # white noise stimuli shown to Retina 01, 02 and 03
│       └── all_images.npy
└── repeating_stimuli_2
        └── all_images.npy # white noise stimuli shown to Retina 04

Unique characteristics

These datasets were collected to investigate how spatial contrast (variations in light intensity) inside receptive fields affects the encoding of visual stimuli by retinal ganglion cells. It specifically aims to address the limitations of linear receptive field models by providing data suitable for testing the Spatial Contrast (SC) model under both artificial and naturalistic conditions.

How is it different from other datasets?

Primate Model: The data is recorded from the marmoset (Callithrix jacchus) retina, offering insights into primate visual processing that differs from common salamander or mouse models.

Fixation Simulation: The natural movie stimuli in this dataset include simulated eye movements (saccades, fixations) modeled based on marmoset viewing statistics.

What features of it make it nice for modelling?

Prediction of responses to dynamic natural stimuli with active vision statistics The naturalistic stimuli in this dataset are overlaid with simulated eye movements (saccades, fixations, and drift) specific to the marmoset visual system. This allows modelers to test how well architectures capture responses to the dynamic spatiotemporal statistics. The high temporal resolution (85 Hz) and fine spatial resolution (7.5 µm/pixel in the case of natural movie data) enable the modeling of precise spike timing and fine spatial integration features.
Benchmarking model generalization and adaptation The dataset’s structure containing paired recordings of the same cells responding to both white noise and natural movies, makes it a great dataset for testing out-of-domain model generalization.
Existing Benchmarks: Extensive benchmarking has already been performed on this data, systematically comparing Linear-Nonlinear (LN) models with various regularizers against Convolutional Neural Networks (CNNs) of varying depths (Vystrčilová et al. 2025) including their generalization capabilities.
Open-Retina Integration: These specific benchmarking tasks (e.g., training on White Noise vs. Natural Movies) and baseline models (LN vs. CNN) from (Vystrčilová et al. 2025) are integrated into open-retina, providing a simple way to benchmark new architectures against established state-of-the-art performance.
Studying the primate visual system Most open retinal datasets come from salamanders or mice. This dataset offers rare access to primate retinal ganglion cells, which are phylogenetically closer to humans. It includes reliably classified primate-specific cell types, including OFF Midget, OFF Parasol, and ON Parasol cells. This allows modelers to investigate distinct functional roles.

Citation information

Original collecting paper: Sridhar, S., Vystrčilová, M., Khani, M. H., Karamanlis, D., Schreyer, H. M., Ramakrishna, V., Krüppel, S., Zapp, S. J., Mietsch, M., Ecker, A. S., & Gollisch, T. Modeling spatial contrast sensitivity in responses of primate retinal ganglion cells to natural movies. biorxiv, 2025

Original source for dataset: Sridhar, S., Gollisch, T. Dataset - Marmoset retinal ganglion cell responses to naturalistic movies and spatiotemporal white noise. (2025) DOI: https://doi.gin.g-node.org/10.12751/g-node.3dfiti/

HuggingFace mirror

Natural movie dataset
White noise dataset

Manuscripts using data from these datasets:

Sridhar, S., Vystrčilová, M., Khani, M. H., Karamanlis, D., Schreyer, H. M., Ramakrishna, V., Krüppel, S., Zapp, S. J., Mietsch, M., Ecker, A. S., & Gollisch, T. Modeling spatial contrast sensitivity in responses of primate retinal ganglion cells to natural movies. biorxiv, 2025. URL
Vystrčilová, Sridhar, S., Burg, M. F., M., Khani, M. H., Karamanlis, D., Schreyer, H. M., Ramakrishna, V., Krüppel, S., Zapp, S. J., Mietsch, Gollisch, T. & Ecker, A. S. A systematic comparison of predictive models on the retina. biorxiv, 2025. URL
Vystrčilová M., Sridhar, S., Burg, M. F., M., Khani, M. H., Karamanlis, D., Schreyer, H. M., Ramakrishna, V., Krüppel, S., Zapp, S. J., Mietsch, Gollisch, T. & Ecker, A. S. Spatial Adaptation of Primate Retinal Ganglion Cells between Artificial and Natural Stimuli. biorxiv, 2025. URL
Vystrčilová M., Sridhar, S., Burg, M. F., Gollisch, T. & Ecker, A. S. Interpreting convolutional neural networks to study wide-field amacrine cell inhibition in the retina. NeurIPS 2025 UniReps workshop. URL
Nellen, N. S., Turishcheva, P., Vystrčilová M., Sridhar S., Gollisch T., Tolias A. S., Ecker A. S. Learning to cluster neuronal function. NeurIPS 2025. URL