Continual learning with attentive recurrent neural networks for temporal data classification.

Continual learning is an emerging research branch of deep learning, which aims to learn a model for a series of tasks continually without forgetting knowledge obtained from previous tasks. Despite receiving a lot of attention in the research community, temporal-based continual learning techniques are still underutilized. In this paper, we address the problem of temporal-based continual learning by allowing a model to continuously learn on temporal data. To solve the catastrophic forgetting problem of learning temporal data in task incremental scenarios, in this research, we propose a novel method based on attentive recurrent neural networks, called Temporal Teacher Distillation (TTD). TTD solves the catastrophic forgetting problem in an attentive recurrent neural network based on three hypotheses, namely Rotation Hypothesis, Redundant Hypothesis, and Recover Hypothesis. Rotation Hypothesis and Redundant hypotheses could cause the attention shift phenomenon, which degrades the model performance on the learned tasks. Moreover, not considering the Recover Hypothesis increases extra memory usage in continuously training different tasks. Therefore, the proposed TTD based on the above hypotheses complements the inadequacy of the existing methods for temporal-based continual learning. For evaluating the performance of our proposed method in task incremental setting, we use a public dataset, WIreless Sensor Data Mining (WISDM), and a synthetic dataset, Split-QuickDraw-100. According to experimental results, the proposed TTD significantly outperforms state-of-the-art methods by up to 14.6% and 45.1% in terms of accuracy and forgetting measures, respectively. To the best of our knowledge, this is the first work that studies continual learning in real-world incremental categories for temporal data classification with attentive recurrent neural networks and provides the proper application-oriented scenario.