Soft-Label Dataset Distillation and Text Dataset Distillation

TL;DR

This paper introduces a novel dataset distillation method that uses soft labels for images and extends to text data, achieving higher accuracy with fewer samples and enabling multi-class encoding per sample.

Contribution

It proposes a soft-label dataset distillation algorithm for images and texts, improving accuracy and reducing sample requirements compared to prior hard-label methods.

Findings

Soft labels improve accuracy by 2-4% on image tasks.

Fewer distilled images achieve high accuracy, e.g., 96% on MNIST with 10 images.

Text distillation outperforms existing methods, reaching near-original accuracy with fewer sentences.

Abstract

Dataset distillation is a method for reducing dataset sizes by learning a small number of synthetic samples containing all the information of a large dataset. This has several benefits like speeding up model training, reducing energy consumption, and reducing required storage space. Currently, each synthetic sample is assigned a single `hard' label, and also, dataset distillation can currently only be used with image data. We propose to simultaneously distill both images and their labels, thus assigning each synthetic sample a `soft' label (a distribution of labels). Our algorithm increases accuracy by 2-4% over the original algorithm for several image classification tasks. Using `soft' labels also enables distilled datasets to consist of fewer samples than there are classes as each sample can encode information for multiple classes. For example, training a LeNet model with 10 distilled images (one per class) results in over 96% accuracy on MNIST, and almost 92% accuracy when trained on just 5 distilled images. We also extend the dataset distillation algorithm to distill sequential datasets including texts. We demonstrate that text distillation outperforms other methods across multiple datasets. For example, models attain almost their original accuracy on the IMDB sentiment analysis task using just 20 distilled sentences. Our code can be found at $\href{https://github.com/ilia10000/dataset-distillation}{\text{https://github.com/ilia10000/dataset-distillation}}$.