N-MNIST Dataset

Neuromorphic MNIST (N-MNIST) [8] is a spiking version of the classic MNIST handwritten digit dataset, created by displaying each MNIST image on a monitor and recording the output of a neuromorphic sensor undergoing saccadic motion.

Overview

Property	Value
Resolution	34 x 34 pixels
Classes	10 digits (0--9)
Training samples	60,000
Test samples	10,000
File format	Binary event files
Event fields	x, y, polarity, timestamp

Usage

from spikeseg.data import NMNISTDataset

train_ds = NMNISTDataset(
    root="/path/to/N-MNIST",
    train=True,
    n_timesteps=10,
    height=34,
    width=34,
    normalize=True,
    polarity_channels=True,
)

for events, label in train_ds:
    # events: (T, 2, 34, 34) voxel grid
    # label: int 0-9
    pass

Purpose in SpikeSEG

N-MNIST serves as a benchmarking dataset. The architecture and STDP training pipeline were originally developed and validated on N-MNIST [1] before being applied to the EBSSA space domain dataset [4].

Citation

G. Orchard, A. Jayawant, G. K. Cohen, and N. Thakor, "Converting static image datasets to spiking neuromorphic datasets using saccades," Frontiers in Neuroscience, vol. 9, p. 437, 2015.