New Directions in Robust Time-Series Machine Learning: Theory, Algorithms, and Applications

Taha Belkhouja

doi:10.7273/000006567

Despite the rapid progress in research on the robustness of deep neural networks (DNNs) for images and text, there is little principled work for the time-series domain. Since time-series data arises in diverse applications, including mobile health, finance, and smart grid, it is important to verify and improve the robustness of DNNs for the time-series domain. Safe deployment of time-series DNNs for real-world applications relies on their ability to be resilient against natural/adversarial perturbations and anomalous inputs that may affect their predictive performance. This dissertation studies the design of robust machine learning (ML) algorithms that aim to minimize both the risk and uncertainty of wrongful decisions made by time-series-based ML systems from both theoretical and algorithmic perspectives. First, we investigate the robustness against adversarial time-series inputs. Adversarial examples were shown to be successful in exposing fundamental blind spots in ML models. While adversarial examples expose how to break the models, the process of creating adversarial examples can itself improve the robustness of ML models by adding them to the training set. The time-series modality poses unique challenges for studying adversarial robustness that are not seen in images and text. The key challenge is how to assess the similarity in the time-series input space to efficiently create valid time-series adversarial examples. Second, we investigate the challenge of Out-of-Distribution (OOD) detection, where the ML system is required to identify time-series inputs that do not follow the distribution of training data. This is a critical task as deep models often make predictions that are very confident yet incorrect on such examples. Detecting OOD examples is challenging, and the potential risks are high for sensitive applications. The key challenge for time-series inputs is how to identify the features that improve the separability between OOD examples and training examples. Motivated by these goals, this dissertation proposes and evaluates a suite of novel solutions to push the frontiers of robust time-series ML: 1) The practical threats of adversarial examples to time-series ML systems; 2) The use of constraints on statistical features of the time-series data to construct adversarial examples, and providing formal robustness certificates for time-series data; 3) The use of elastic measures such as Dynamic Time Warping to quantify the similarity between time-series examples and developing theoretically-sound algorithms to efficiently construct valid adversarial examples, and to train robust ML models by explicitly solving a min-max optimization problem; 4) Adapting and applying the developed algorithms to real-world applications including wearable sensors enabled ML systems for healthcare to handle both natural perturbations and missing sensor data; and 5) A novel OOD detection algorithm based on deep generative models for the time-series domain and explain why prior OOD methods from the other domains perform poorly.

New Directions in Robust Time-Series Machine Learning: Theory, Algorithms, and Applications

Files and links (1)

Abstract

Metrics

Details