Multidigitmnist: 100 and 1,000 Class MNIST Variants for Meta-Learning

Multidigitmnist provides synthetic image data generated by combining multiple MNIST digits to create classification tasks with 100 or 1,000 distinct classes. Created by shaohua0116 and updated in 2021, the project facilitates meta-learning research by expanding the label space of the original MNIST digits.

Use Cases

• Benchmarking meta-learning algorithms using the 1,000-class label set
• Training few-shot learning models on synthetic multi-digit image sequences
• Evaluating classification accuracy across expanded 100-class digit combinations

Strengths

• Provides 1,000 distinct class labels for high-cardinality testing
• Released under the MIT license for unrestricted research use
• Specifically formatted for meta-learning tasks unlike standard MNIST

Limitations

• Synthetic digit combinations do not capture the complexity of real-world handwritten strings
• MNIST-based data is often considered a 'toy' problem in modern computer vision

Provenance

Source

shaohua0116 GitHub repository

Collection Method

synthetic

Freshness

Last updated July 2021.