An Empirical Analysis of the Impact of PRNG and External Seeds on Performance and Prediction Diversity in Deep Learning Models

This study empirically analyzes how seed types affect the performance and prediction diversity of deep learning models. The experiments used a consistent CNN architecture and the CIFAR-10 dataset, with 30 repeated training runs conducted under two settings: fixed seeds based on pseudorandom number generators (PRNGs) and non-deterministic external seeds. The results showed that models using external seeds achieved a slightly higher average accuracy of 68.3%, compared to 68.0% for PRNG-based models, and demonstrated more stable overall convergence behavior. Analysis of convergence trends revealed that the final loss was 0.0711 for external seeds and 0.0795 for PRNG seeds, with standard deviations of 0.0134 and 0.0427, respectively, indicating faster and more consistent training. In the weight distribution analysis, the external seed condition showed broader and more balanced patterns, suggesting enhanced exploration in the parameter space. The average L2 distance of softmax outputs was 70.068 for external seeds and 70.33 for PRNG seeds, while the standard deviation was 0.794 and 1.569, respectively, indicating greater consistency in predictions. These results demonstrate that seed configuration plays a critical role in ensuring reproducibility and diversity in deep learning experiments, beyond simply setting initial values.