# InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning

**Authors:** Ziheng Qin, Kai Wang, Zangwei Zheng, Jianyang Gu, Xiangyu Peng, Daquan Zhou, Yang You  
**arXiv:** [2303.04947](https://arxiv.org/abs/2303.04947)  
**Date:** 2023-03-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.04947v1/

## Key points

- InfoBatch is a framework that accelerates lossless training through unbiased dynamic data pruning.
- It randomly prunes based on the loss distribution and rescales gradients of remaining samples to mitigate bias.
- Training the full dataset in the last few epochs further enhances performance and reduces total update bias.
- InfoBatch consistently achieves lossless training results across datasets like CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K with significant cost savings.
- Extension of InfoBatch into semantic segmentation tasks shows lossless mIoU on ADE20K dataset with an overall cost reduction.
- InfoBatch accelerates in data dimensionality, speeding up large-batch training methods without compromising performance or extra costs.
- The authors plan to make their code public for broader adoption in machine learning applications.

## Summary

, , , , 

The paper "InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning" presents InfoBatch, a novel framework that achieves lossless training acceleration through unbiased dynamic data pruning. Traditional methods for data pruning often lead to gradient expectation bias between pruned and original data by filtering out less informative samples. In contrast, InfoBatch randomly prunes based on the loss distribution and rescales gradients of remaining samples to mitigate this bias. The authors also propose training the full dataset in the last few epochs to further enhance performance and reduce total update bias. As a plug-and-play and architecture-agnostic framework, InfoBatch consistently achieves lossless training results across various datasets such as CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K with cost savings of 40%, 33%, 30%, and 26% respectively. The extension of InfoBatch into semantic segmentation tasks also demonstrates lossless mean Intersection over Union (mIoU) on the ADE20K dataset with a 20% overall cost reduction. Additionally, InfoBatch accelerates in data dimensionality, speeding up large-batch training methods like LARS and LAMB by 1.3 times without compromising performance or incurring extra costs. The authors plan to make their code public for broader adoption and implementation in various machine learning applications.