Natural Scenes Dataset (NSD)

The Natural Scenes Dataset (NSD) is a large-scale 7 Tesla fMRI dataset acquired from 8 human participants viewing tens of thousands of natural scene images (Allen et al. 2022). It has become a foundational resource in visual neuroscience, bridging cognitive neuroscience and artificial intelligence by providing high-quality fMRI responses to richly varied naturalistic stimuli.

Overview

• Species: Human (Homo sapiens)
• Modality: 7T fMRI (whole-brain, high-resolution)
• Stimuli: ~73,000 natural scene images from the Microsoft COCO dataset (Lin et al. 2014)
• Participants: 8 healthy adults, each scanned across 30–40 sessions
• Published: Allen et al. (2022), Nature Neuroscience

Key Contributions

1. Large-scale, high-quality fMRI data — single-trial response estimates using GLMsingle (Prince et al. 2022), enabling robust encoding model training
2. Bridge to AI — enabled training of neural network encoding models that predict voxel-wise responses to arbitrary natural images
3. Cross-species extension — the same stimulus paradigm (subset of 1,000 images) was extended to macaques in the Triple-N dataset

Relationship to the Triple-N Dataset

The NSD provides the human fMRI counterpart to the Triple-N macaque dataset. Together they enable:

• Direct cross-species comparison using identical natural scene stimuli
• Linking single-neuron macaque data (Triple-N Neuropixels) to human fMRI voxel responses (NSD)
• Validation of fMRI-defined category-selective regions against ground-truth single-unit recordings in macaques
• Representational geometry alignment between human and macaque visual cortex

Significance

The NSD has become a benchmark for:

• fMRI encoding/decoding models of natural vision
• Training and evaluating artificial neural networks as models of the visual system
• Cross-species comparative neuroscience (in conjunction with the Triple-N dataset)
• Understanding the organization of high-level visual cortex under naturalistic conditions

References

• Allen, E.J. et al. (2022). A massive 7T fMRI dataset to bridge cognitive neuroscience and artificial intelligence. Nat. Neurosci. 25, 116–126. 10.1038/s41593-021-00962-x
• Prince, J.S. et al. (2022). Improving the accuracy of single-trial fMRI response estimates using GLMsingle. eLife 11, e77599. 10.7554/eLife.77599
• Li, Y. et al. (2026). Triple-N dataset: large-scale fMRI-guided dense recordings of nonhuman primate neural responses to natural scenes. Nature Neuroscience. 10.1038/s41593-026-02322-z