On Model Stability as a Function of Random Seed

TL;DR

This paper investigates how random seed choices affect model stability and interpretability, and introduces techniques to enhance robustness and consistency across different random initializations.

Contribution

It presents ASWA and NASWA methods that significantly improve model stability against random seed variations, reducing performance variability.

Findings

Random seeds can cause significant variability in model interpretations.

ASWA and NASWA techniques improve model robustness and reduce performance variance.

On average, these methods reduce the standard deviation of model performance by 72%.

Abstract

In this paper, we focus on quantifying model stability as a function of random seed by investigating the effects of the induced randomness on model performance and the robustness of the model in general. We specifically perform a controlled study on the effect of random seeds on the behaviour of attention, gradient-based and surrogate model based (LIME) interpretations. Our analysis suggests that random seeds can adversely affect the consistency of models resulting in counterfactual interpretations. We propose a technique called Aggressive Stochastic Weight Averaging (ASWA)and an extension called Norm-filtered Aggressive Stochastic Weight Averaging (NASWA) which improves the stability of models over random seeds. With our ASWA and NASWA based optimization, we are able to improve the robustness of the original model, on average reducing the standard deviation of the model's performance by 72%.