Reproduce “On Model Stability as a Function of Random Seed” on OpenResearch

About this paper

Pranava Madhyastha, Rishabh Jain

This study quantifies the impact of random seeds on deep neural network performance and interpretation consistency, revealing substantial variability in standard models. It introduces Aggressive Stochastic Weight Averaging (ASWA) and Norm-filtered ASWA (NASWA) methods, which notably reduce performance standard deviation by an average of 72% and improve interpretation stability (e.g., attention entropy by approximately 60%) across diverse NLP tasks.

Key insights

• Model interpretations are highly dependent on the specific random seed used during training, often leading to divergent explanations for functionally similar models.
• Standard deep neural networks show significant variability in both prediction performance and the consistency of attention, gradient-based, and LIME interpretations.
• Optimizing with aggressive weight averaging methods (ASWA/NASWA) can stabilize training, promote convergence to wider minima, and yield more robust models with consistent interpretations.

Summary by alphaXiv.