Deep Learning on a Data Diet: Finding Important Examples Early in Training

TL;DR

This paper introduces simple early-in-training scores, GraNd and EL2N, to identify important training examples in vision datasets, enabling data pruning and improved generalization with minimal training data.

Contribution

The paper proposes two novel early-in-training scores, GraNd and EL2N, for identifying important examples, allowing effective data pruning and insights into training dynamics.

Findings

EL2N scores can prune half of CIFAR10 data with slight accuracy improvement.

Scores generalize across architectures and hyperparameters.

Pruning based on these scores does not sacrifice test accuracy.

Abstract

Recent success in deep learning has partially been driven by training increasingly overparametrized networks on ever larger datasets. It is therefore natural to ask: how much of the data is superfluous, which examples are important for generalization, and how do we find them? In this work, we make the striking observation that, in standard vision datasets, simple scores averaged over several weight initializations can be used to identify important examples very early in training. We propose two such scores -- the Gradient Normed (GraNd) and the Error L2-Norm (EL2N) scores -- and demonstrate their efficacy on a range of architectures and datasets by pruning significant fractions of training data without sacrificing test accuracy. In fact, using EL2N scores calculated a few epochs into training, we can prune half of the CIFAR10 training set while slightly improving test accuracy. Furthermore, for a given dataset, EL2N scores from one architecture or hyperparameter configuration generalize to other configurations. Compared to recent work that prunes data by discarding examples that are rarely forgotten over the course of training, our scores use only local information early in training. We also use our scores to detect noisy examples and study training dynamics through the lens of important examples -- we investigate how the data distribution shapes the loss surface and identify subspaces of the model's data representation that are relatively stable over training.