GRAFT: Gradient-Aware Fast MaxVol Technique for Dynamic Data Sampling

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Jha, Ashish, Phan, Anh huy, Dibo, Razan, Leplat, Valentin
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912547995648000
author Jha, Ashish
Phan, Anh huy
Dibo, Razan
Leplat, Valentin
author_facet Jha, Ashish
Phan, Anh huy
Dibo, Razan
Leplat, Valentin
contents Training modern neural networks on large datasets is computationally and environmentally costly. We introduce GRAFT, a scalable in-training subset selection method that (i) extracts a low-rank feature representation for each batch, (ii) applies a Fast MaxVol sampler to select a small, diverse subset that spans the batch's dominant subspace, and (iii) dynamically adjusts the subset size using a gradient-approximation criterion. By operating in low-rank subspaces and training on carefully chosen examples instead of full batches, GRAFT preserves the training trajectory while reducing wall-clock time, energy consumption, and $\mathrm{CO}_2$ emissions. Across multiple benchmarks, GRAFT matches or exceeds recent selection baselines in both accuracy and efficiency, providing a favorable trade-off between accuracy, efficiency, and emissions.
format Preprint
id arxiv_https___arxiv_org_abs_2508_13653
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle GRAFT: Gradient-Aware Fast MaxVol Technique for Dynamic Data Sampling
Jha, Ashish
Phan, Anh huy
Dibo, Razan
Leplat, Valentin
Machine Learning
Artificial Intelligence
Numerical Analysis
Training modern neural networks on large datasets is computationally and environmentally costly. We introduce GRAFT, a scalable in-training subset selection method that (i) extracts a low-rank feature representation for each batch, (ii) applies a Fast MaxVol sampler to select a small, diverse subset that spans the batch's dominant subspace, and (iii) dynamically adjusts the subset size using a gradient-approximation criterion. By operating in low-rank subspaces and training on carefully chosen examples instead of full batches, GRAFT preserves the training trajectory while reducing wall-clock time, energy consumption, and $\mathrm{CO}_2$ emissions. Across multiple benchmarks, GRAFT matches or exceeds recent selection baselines in both accuracy and efficiency, providing a favorable trade-off between accuracy, efficiency, and emissions.
title GRAFT: Gradient-Aware Fast MaxVol Technique for Dynamic Data Sampling
topic Machine Learning
Artificial Intelligence
Numerical Analysis
url https://arxiv.org/abs/2508.13653