Underwater Acoustic Data Classification Using Continual Learning

Underwater acoustic data classification has received significant attention from the research community in recent years because of its potential applications in underwater object detection and classification. Underwater acoustic data classification presents a challenging problem due to several factors, including the complicated sound waves propagation in marine environment, diminished illumination, frequency-dependent absorption, dispersion of light, and the presence of complex background noise. Underwater sound propagation is largely effected by underwater environment and geographical change that is sound of same vessel is observed with different characteristics. The artificial intelligence based system may not perform well to classify the same data in different sea environments because model trained in one environment data may not perform well in other environment data. There is a need to make AI based system capable to learn in different sea environmental conditions and geographical location. In underwater acoustic classification field, it is very important for artificial intelligence based models to be able to learn and adapt across different tasks. To fill this gap, continual learning techniques are applied on the underwater acoustic dataset in order to improve classification accuracy. The proposed model is evaluated on benchmark real world dataset and the results demonstrates the superiority of the proposed model. Two famous techniques of continual learning are implemented in this proposed model which are experience replay and elastic weight consolidation. Experience replay techniques shows perfect balance between stability (not forgetting old knowledge) and plasticity (ablility to learn new tasks) as compare to elastic weight consolidation. Elastic weight consolidation also shown significant improved classification accuracy.