ADV-ResNet: Residual Network with Controlled Adversarial Regularization for Effective Classification of Practical Time Series Under Training Data Scarcity Problem

Practical time series datasets in classification tasks often suffer from scarcity in number of training instances owing to the expenses associated with the annotation exercise. Deep learning algorithms generally demand sufficiency in the seen examples for its learning purposes. In this paper, we consider adversarial perturbation as the set of invariants such that a robust model can be constructed under the practical constraints of training sample scarcity. We propose ADV-ResNet that augments the learning ability of a Residual Network (ResNet) through adversarial regularization, where adversarial training is transformed to robust form of data augmentation. We propose novel algorithm for adversarial regularization factor computation that estimates the amount of regularization for controlled perturbation in the learning process. The intentional yet measured introduction of perturbations in the ResNet training process enables it to learn better under crafted and controlled perturbations that constitute unseen but important input space. Our empirical investigation on publicly available time series classification datasets (UCR time series archive) demonstrates the utility of ADV-ResNet through ablation study. The empirical evidence clearly indicates the superiority of ADV-ResNet over the baseline ResNet, as well as ADV-ResNet significantly outperforms the state-of-the-art time series classification models.