On Model Stability as a Function of Random Seed

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 929–939 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 929 On Model Stability as a Function of Random Seed Pranava Madhyastha Department of Computing Imperial College London pranava@imperial.ac.uk Rishabh Jain∗ Bloomberg London rjain213@bloomberg.net Abstract In this paper, we focus on quantifying model stability as a function of random seed by in- vestigating the effects of the induced random- ness on model performance and the robustness of the model in general. We specifically per- form a controlled study on the effect of random seeds on the behaviour of attention, gradient- based and surrogate model based (LIME) in- terpretations. Our analysis suggests that ran- dom seeds can adversely affect the consistency of models resulting in counterfactual interpre- tations. We propose a technique called Aggres- sive 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 ro- bustness of the original model, on average re- ducing the standard deviation of the model’s performance by 72%. 1 Introduction There has been a tremendous growth in deep neu- ral network based models that achieve state-of- the-art performance. In fact, most recent end- to-end deep learning models have surpassed the performance of careful human feature-engineering based models in a variety of NLP tasks. However, deep neural network based models are often brit- tle to various sources of randomness in the training of the models. This could be attributed to several sources including, but not limited to, random pa- rameter initialization, random sampling of exam- ples during training and random dropping of neu- rons. It has been observed that these models have, more often, a set of random seeds that yield better results than others. This has also lead to research ∗This work was conducted when the author was a student at Imperial College London. suggesting random seeds as an additional hyper- parameter for tuning (Bengio, 2012)1. One possi- ble explanation for this behavior could be the exis- tence of multiple local minima in the loss surface. This is especially problematic as the loss surfaces are generally non-convex and may have multiple saddle points making it difficult to achieve model stability. if high crimes were any more generic it would have a universal product code instead of a title (Pr (Ynegative) = 0.99) if high crimes were any more generic it would have a universal product code instead of a title (Pr (Ynegative) = 0.98) Figure 1: Importance based on attention probabil- ities for two runs of the same model with same parameters and same hyperparameters, but with two different random seeds (color magnitudes: pink<magenta<red) Recently the NLP community has witnessed a resurgence in interpreting and explaining deep neural network based models (Jain et al., 2019; Jain and Wallace, 2019; Alvarez-Melis and Jaakkola, 2017). Most of the interpretation based methods involve one of the following ways of …

Content truncated to 500 words