TAROT: Targeted Data Selection via Optimal Transport

TL;DR

TAROT introduces a novel data selection method based on optimal transport theory that effectively handles complex, multimodal data distributions, outperforming existing influence-based heuristics across various deep learning tasks.

Contribution

The paper presents TAROT, a new targeted data selection framework that uses whitened feature distance and optimal transport to improve selection in multimodal data scenarios.

Findings

TAROT outperforms state-of-the-art methods in multiple tasks.

It effectively handles multimodal and complex data distributions.

The method provides reliable estimates of optimal selection ratios.

Abstract

We propose TAROT, a targeted data selection framework grounded in optimal transport theory. Previous targeted data selection methods primarily rely on influence-based greedy heuristics to enhance domain-specific performance. While effective on limited, unimodal data (i.e., data following a single pattern), these methods struggle as target data complexity increases. Specifically, in multimodal distributions, these heuristics fail to account for multiple inherent patterns, leading to suboptimal data selection. This work identifies two primary factors contributing to this limitation: (i) the disproportionate impact of dominant feature components in high-dimensional influence estimation, and (ii) the restrictive linear additive assumptions inherent in greedy selection strategies. To address these challenges, TAROT incorporates whitened feature distance to mitigate dominant feature bias, providing a more reliable measure of data influence. Building on this, TAROT uses whitened feature distance to quantify and minimize the optimal transport distance between the selected data and target domains. Notably, this minimization also facilitates the estimation of optimal selection ratios. We evaluate TAROT across multiple tasks, including semantic segmentation, motion prediction, and instruction tuning. Results consistently show that TAROT outperforms state-of-the-art methods, highlighting its versatility across various deep learning tasks. Code is available at https://github.com/vita-epfl/TAROT.