Data augmentation using synthetic data for time series classification\n with deep residual networks

Abstract

Data augmentation in deep neural networks is the process of generating\nartificial data in order to reduce the variance of the classifier with the goal\nto reduce the number of errors. This idea has been shown to improve deep neural\nnetwork's generalization capabilities in many computer vision tasks such as\nimage recognition and object localization. Apart from these applications, deep\nConvolutional Neural Networks (CNNs) have also recently gained popularity in\nthe Time Series Classification (TSC) community. However, unlike in image\nrecognition problems, data augmentation techniques have not yet been\ninvestigated thoroughly for the TSC task. This is surprising as the accuracy of\ndeep learning models for TSC could potentially be improved, especially for\nsmall datasets that exhibit overfitting, when a data augmentation method is\nadopted. In this paper, we fill this gap by investigating the application of a\nrecently proposed data augmentation technique based on the Dynamic Time Warping\ndistance, for a deep learning model for TSC. To evaluate the potential of\naugmenting the training set, we performed extensive experiments using the UCR\nTSC benchmark. Our preliminary experiments reveal that data augmentation can\ndrastically increase deep CNN's accuracy on some datasets and significantly\nimprove the deep model's accuracy when the method is used in an ensemble\napproach.\n