FSTS: A Feature Space Transfer Selection Method for Data

Jizhe Liu, Yifan Zhu, Yunjun Gao

Keywords: Core Dataset Selection, Feature Space Transfer, Data Exchange

TL;DR: We propose a feature-space transfer method to select core subsets from target datasets using source dataset criteria

Abstract: The performance of deep learning image models is often largely determined by data quality. However, high-quality data is often scarce and difficult to collect. The data exchange platform provides a promising solution for obtaining image training samples, but the actual data exchange must consider the costs associated with data collection, storage, and training. This article proposes a Feature Space Transfer Selection (FSTS) method for identifying core data subsets that are crucial for model training. The proposed method extracts feature vectors from the source and target datasets, calculates class centroids from the reference (source) dataset, and ranks the target samples based on their similarity to these class centroids. Then select the core dataset from the target data based on the similarity ranking. The experimental results show that FSTS outperforms prior state-of-the-art approaches and effectively helps users select the core set of training data, which helps improve the efficiency of model training and overall performance.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 11304