A Sequential Experience Network Based Continual Learning for Air Crisis Event Recognition

With the exponential growth of multimodal data on the Internet, the recognition of air crisis events has become increasingly vital for managing accident-related information. This work underscores the continuous emergence of new air crisis events as its central feature. Traditional studies in this field have been limited to identifying a predefined array of these events. Nevertheless, the real-world application requires the dynamic recognition and management of unforeseen air crisis events. A static recognition model proves inadequate in such a rapidly changing environment, as it fails to identify emerging events. To address this deficiency, we introduce a novel problem termed Continual Air Crisis Event Recognition (CACER) in this paper. CACER necessitates a recognition model capable of learning from sequentially collected training data and categorizing all acquired events in test data. We propose a new Sequential Experience Network (SEN) designed to learn continuously from sequential data and recognize previously unidentified air crisis events. Our approach begins with current experience learning, aimed at mastering the classification of newly emerged events within current datasets. Subsequently, we employ a method of prior experience replay enhanced by self-knowledge distillation to solidify previously learned knowledge and avert catastrophic forgetting. In anticipation of entirely new events, we also intro-duce an unforeseen experience preparation component, featuring a modality mixture mechanism to prime classifiers. Furthermore, we conduct extensive experiments on the AirCrisisMMD and CrisisMMD datasets. The empirical results across these two datasets assert the superiority of our method over leading-edge alternatives.