Adaptive Enhancement of an Active Sonar Classifier Using Mode-Connectivity-Based Fine-Tuning Under Data Set Shifts

In supervised-learning-based active sonar classification overcoming data set shifts through standard fine-tuning is challenging due to the limited size and diversity of active sonar data sets. To address this challenge, we propose a robust fine-tuning method using mode connectivity (RoFT-MC), which mitigates two key problems in standard fine-tuning: catastrophic forgetting and negative transfer. RoFT-MC constructs a mode connectivity curve between two independently pretrained models. For adaptation, the curve parameters are optimized using in situ test data rather than training data. RoFT-MC effectively adapts to the shifted test data set while maintaining its performance on the training data set by ensuring that the fine-tuned weights remain on the curve. In addition, we utilize a feasible fine-tuning data set composed of test clutter samples combined with training target samples instead of unavailable test target samples to avoid biased predictions. In the efficacy examination standard fine-tuning failed to adapt to the shifted test data set, whereas RoFT-MC demonstrated a significant performance improvement. Specifically, RoFT-MC increased the probability of detection from 0.2710 to 0.6438 at a false alarm rate of 0.1, while maintaining comparable performance on the training data set.