Synthetic Condensed Image Data for MNIST SVHN and CIFAR-10

This repository contains small synthetic data for the MNIST, SVHN, and CIFAR-10 image datasets. The data consists of images and corresponding labels, with sizes ranging from 1, 10, to 50 images per class (IPC). It was created by ICML2022 for research on dataset condensation.

Use Cases

• Train models on synthetic images and labels to evaluate the efficacy of dataset condensation techniques.
• Benchmark the performance of condensed data from MNIST, SVHN, and CIFAR-10 against original training sets.
• Analyze the relationship between the number of images per class (IPC) and model accuracy on downstream tasks.

Strengths

• Provides condensed versions of three standard benchmark datasets: MNIST, SVHN, and CIFAR-10.
• Offers multiple data sizes, specifically 1, 10, and 50 images per class (IPC).
• Data is directly associated with a published ICML 2022 research paper.

Limitations

• The dataset is very small, consisting only of synthetic condensed data, not the full original datasets.
• The specific file formats, total row count, and column structure are not provided in the input.
• The synthetic data may not capture the full variance and complexity of the original datasets.

Provenance

Source

ICML2022, associated with the GitHub repository 'snu-mllab/Efficient-Dataset-Condensation'.

Collection Method

Synthetic data generated via dataset condensation techniques as described in the associated research paper.

Freshness

Last updated on 2022-08-01.