DeepNurse: Self-healing Neural Networks for Unknown Domains

Neural networks (NNs) have been widely adopted in a variety of applications, but their ability to work in dynamic and unpredictable environments remains a challenge. NNs are trained with large, yet constrained, datasets that can only cover a limited set of scenarios and features, leaving them vulnerable to failure when employed in unknown domains. Although some techniques have been proposed to enhance generalization of NNs across new domains, the vast and unpredictable nature of real-world environments calls for solutions that can self-heal the NN when facing unknown scenarios. This paper presents DeepNurse , a self-healing solution that aims to automatically adapt NNs to handle covariate shifts, where the input data distribution changes (e.g., different lighting, textures, or sensor noise) while the core relationship between the content and the correct output remains the same. Our solution focuses on NNs that take images as input (vision NNs) and uses uncertainty estimators and changepoint detectors to identify new domains. When an unknown domain is detected, DeepNurse uses neural style transfer to generate a dataset of images that captures its main characteristics. These data are then used to adapt the NN through a fine-tuning process. An extensive empirical evaluation demonstrates that DeepNurse can detect most of the unknown domains and improves the performance of NNs by 13.0% on average.