DRIP: DRop unImportant data Points -- Enhancing Machine Learning Efficiency with Grad-CAM-Based Real-Time Data Prioritization for On-Device Training

2025-04-11

AI Summary

To address the trade-off between excessive storage overhead and performance degradation in edge-device model training under resource constraints, this paper proposes DRIP—a real-time data importance-aware online training framework for embedded systems. Methodologically, DRIP is the first to enable dynamic, on-the-fly retention or discarding of data points without requiring full-dataset access; it introduces a lightweight, Grad-CAM–driven DRIP Score for real-time importance estimation directly on-device; and integrates embedded-aware optimization algorithms with streaming-data filtering. Evaluated on four benchmark datasets, DRIP reduces average storage consumption by 39% while maintaining or surpassing the accuracy of full-dataset training. Its core contributions are: (1) the first end-to-edge real-time data selection paradigm, and (2) an open-source, deployable lightweight importance scoring mechanism tailored for embedded environments.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Systems and Infrastructure for Web, Mobile and WoT: Data management and stream processing for Web, mobile and wireless applications

Abstract

Selecting data points for model training is critical in machine learning. Effective selection methods can reduce the labeling effort, optimize on-device training for embedded systems with limited data storage, and enhance the model performance. This paper introduces a novel algorithm that uses Grad-CAM to make online decisions about retaining or discarding data points. Optimized for embedded devices, the algorithm computes a unique DRIP Score to quantify the importance of each data point. This enables dynamic decision-making on whether a data point should be stored for potential retraining or discarded without compromising model performance. Experimental evaluations on four benchmark datasets demonstrate that our approach can match or even surpass the accuracy of models trained on the entire dataset, all while achieving storage savings of up to 39%. To our knowledge, this is the first algorithm that makes online decisions about data point retention without requiring access to the entire dataset.

Problem

• Prioritize data points for efficient on-device training
• Reduce storage needs without losing model accuracy
• Enable real-time data retention decisions using Grad-CAM

Innovation

• Uses Grad-CAM for real-time data prioritization
• Computes DRIP Score for dynamic data retention
• Optimizes on-device training with storage savings