A study on practical adaptation of active sonar classifiers under dataset shifts

Active sonar classifiers in practical environments face significant challenges due to dataset shifts. While fine-tuning is a widely used technique to address such problems, its application to active sonar classification is hindered by the small size and limited diversity of available datasets, resulting in catastrophic forgetting and negative transfer. In this study, we introduce a mode-connectivity-based fine-tuning to overcome these challenges. This approach constructs a mode-connectivity curve between the weights of two independently pre-trained networks and optimizes the curve parameters using in-situ test data instead of offline training data. With the guidance of the mode-connectivity curve, this approach ensures that the fine-tuned weights remain robust to both training and test datasets, enabling effective adaptation to shifted test datasets while maintaining performance on the training dataset. Additionally, we extend this method to various architectures and datasets. Experimental results demonstrate its robustness across diverse conditions, highlighting its potential applicability to a wide range of applications.