A Coreset Selection Framework with Ensemble Aggregation for Image Classification

2026-07-10 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

A new coreset selection framework, combining SCOre-Stratified Selection (SCOSS) with ensemble aggregation, addresses the challenges of training models on large-scale image datasets by reducing time and memory costs. SCOSS partitions training data into score-based intervals and samples from each, while the ensemble aggregates predictions from multiple runs on independently sampled subsets. This framework was evaluated against moderate and random selection baselines, both in original and class-balanced versions, using Simple Graph Convolution (SGC) and Support Vector Machine (SVM) classifiers. Experiments demonstrate that SCOSS is competitive, frequently performing best for SGC, and facilitates favorable trade-offs between accuracy and efficiency. Notably, on fine-grained datasets, SGC utilizing SCOSS surpasses SVMs even when employing fewer labeled samples. The associated code and supplementary materials are publicly available.

Key takeaway

For Machine Learning Engineers developing image classification models with large datasets, consider implementing the SCOre-Stratified Selection (SCOSS) framework. This approach can significantly reduce training time and memory costs while maintaining or improving accuracy, especially with Simple Graph Convolution (SGC) classifiers. You should evaluate SCOSS to achieve favorable trade-offs between efficiency and performance, potentially outperforming traditional methods like SVMs with fewer labeled samples.

Key insights

The framework combines coreset selection and ensemble aggregation to efficiently train image classifiers on large datasets.

Principles

• Representative subsets reduce training costs.
• Ensemble aggregation improves prediction robustness.
• Stratified sampling enhances coreset diversity.

Method

Proposes SCOre-Stratified Selection (SCOSS) to partition training data into score-based intervals and sample from each. Aggregates predictions from multiple runs on independently sampled subsets.

In practice

• Use SCOSS for SGC classifiers.
• Apply ensemble aggregation for robustness.
• Explore stratified sampling for large datasets.

Topics

• Coreset Selection
• Ensemble Aggregation
• Image Classification
• SCOre-Stratified Selection
• Simple Graph Convolution
• Large-scale Datasets

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Graph Neural Networks for Semi-Supervised Image Classification with Multi-Feature Aggregation — Integrating diverse feature and graph representations with GNNs significantly boosts semi-supervised image classification accuracy.
• Simplex-Constrained Sparse Bagging: Transitioning from Uniform Priors to Sparse Posteriors in Ensemble Learning — SCSB compresses and calibrates bagging ensembles by optimizing sparse voting weights, improving efficiency and confidence.
• Few-Medoids: An Embarrassingly Simple Coreset Selection Method for Few-Shot Knowledge Distillation — Few-medoids, a simple geometric centrality method in the teacher's latent space, consistently outperforms complex coreset selection for few-shot knowledge distillation.
• Selectivity Drives Efficiency: Dataset Pruning for Visual Place Recognition — Place-wise dataset pruning, using intra-place diversity and inter-place similarity, significantly enhances visual place recognition data efficiency.
• Decomposing one-class support vector machine into an ensemble of one-data support vector machines — Decomposing OCSVM into an ensemble of one-data models significantly accelerates processing for large datasets.
• Beyond the Performance Illusion: Structure-Aware Stratified Partitioning and Curriculum Distributionally Robust Optimization for Spatially Correlated Domains — Conventional random dataset splits fail in correlated domains, necessitating new evaluation and training methods.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.