Optimizing Dataset Quality and Diversity in Neural Networks: A Study of the Vendi Score

We investigate methods for evaluating and optimizing dataset quality in the context of image-based object detection tasks, focusing on the Vendi Score (VS) and its enhanced variant, the quality-weighted Vendi Score (qVS). These metrics provide a robust framework for assessing and improving dataset diversity, a critical factor in effectively training neural networks. In particular, the qVS allows for a customizable quality measure to account for dataset-specific characteristics. In this paper, we introduce a specific weighting function based on the ratio of objects to total images to ensure balanced dataset composition and to emphasize the importance of diverse object representations. Feature extraction techniques are employed to represent image objects in a unified feature space, enabling similarity calculations and diversity assessments. Additionally, a multi objective genetic algorithm (MOGA) is utilized to optimize datasets across multiple classes, maximizing diversity while maintaining class balance. Our experimental results demonstrate that models trained on subsets optimized for diversity using the qVS achieve improved performance. Diverse subsets lead to higher precision and generalization capabilities compared to randomly selected or less diverse datasets of the same size. Especially when sampling minibatches during training this method may prove beneficial, as it allows for a more representative sample of the dataset. These findings underscore the pivotal role of dataset diversity in enhancing neural network performance and highlight the utility of the VS and qVS as valuable tools for strategically shaping dataset composition to improve outcomes.

2025 10th International Conference on Frontiers of Signal Processing (ICFSP)