Optimizing Object Detection Model Training Using ChatGPT for Dataset Reduction
Aleksa Stokanović, Milena Milošević, Željko Lukač · 2025
Artificial intelligence (AI) has significantly advanced in recent years, with deep neural networks playing a crucial role in various applications, including object detection. Training deep learning models for object detection requires large amounts of data to ensure high accuracy and generalization. However, handling such extensive datasets introduces challenges in computational efficiency and resource utilization. Although large datasets generally improve generalization, not all data points contribute equally to model performance. Redundant, low-quality, or overly simple samples can increase training time and resource consumption while providing little benefit. Moreover, using more data often leads to diminishing returns, longer convergence times, and computational inefficiency. Therefore, selective data reduction presents an opportunity to improve efficiency without compromising accuracy. Optimizing dataset size while maintaining performance is essential for improving the efficiency of object detection models. Traditional methods for data reduction often require manual intervention or predefined heuristics, making them time-consuming and resource-intensive. This paper investigates the effectiveness of ChatGPT as a tool for automated dataset reduction (by selecting high-value data points—those that contribute the most to model learning), aiming to address the problem of long training times in deep neural networks. By evaluating different approaches, including step-by-step guidance, fully automated pruning, and a specialized Data Analyst GPT model, this study assesses how well ChatGPT can optimize datasets for object detection training (automotive use-cases) while preserving model performance.