Training Robust Deep Models for Time-Series Domain: Novel Algorithms and Theoretical Analysis

Despite the success of deep neural networks (DNNs) for real-world applications over time-series data such as mobile health, little is known about how to train robust DNNs domain due its unique characteristics compared images and text data. In this paper, we fill gap by proposing a novel algorithmic framework referred RObust Training Time-Series (RO-TS) create models classification tasks. Specifically, formulate min-max optimization problem model parameters explicitly reasoning robustness criteria in terms additive perturbations inputs measured global alignment kernel (GAK) based distance. We also show generality advantages our formulation using summation structure alignments relating both GAK dynamic time warping (DTW). This an instance family compositional problems, which are challenging open with unclear theoretical guarantee. propose principled stochastic alternating gradient descent ascent (SCAGDA) algorithm problems. Unlike traditional methods that require approximate computation distance measures, SCAGDA approximates on-the-fly moving average approach. theoretically analyze convergence rate provide strong support estimation Our experiments on benchmarks demonstrate RO-TS creates more when adversarial training prior rely augmentation or new definitions loss functions. importance Euclidean