MNIST: 70,000 Grayscale Handwritten Digit Images for Classification

MNIST contains 70,000 28x28 grayscale images of handwritten digits (0-9) sourced from NIST databases by Yann LeCun. The collection is split into 60,000 training and 10,000 validation samples, with exactly 7,000 images per digit class.

Use Cases

• Training multi-class classifiers to identify digit labels 0-9
• Benchmarking neural network architectures on 28x28 pixel intensity features
• Evaluating clustering algorithms on grayscale image data

Strengths

• 70,000 total images
• Perfectly balanced with 7,000 images per digit class
• Expert-generated annotations from NIST databases

Limitations

• Low 28x28 resolution limits use for high-fidelity OCR tasks
• Historical demographic bias from 1990s Census employees and students

Provenance

Source

NIST (National Institute of Standards and Technology)

Collection Method

Handwritten samples collected from Census Bureau employees and high school students

Freshness

Last updated August 2024

Geography

United States