Reproduce “Cross-Domain Knowledge Transfer for Underwater Acoustic Classification Using Pre-trained Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Amirmohammad Mohammadi, Tejashri Kelhe, Davelle Carreiro, Alexandra Van Dine, Joshua Peeples
Researchers from Texas A&M and MIT Lincoln Lab evaluated cross-domain knowledge transfer for underwater acoustic classification, finding that models pre-trained on large image datasets like ImageNet can achieve comparable or superior performance to those pre-trained on audio datasets for the DeepShip dataset. The study also investigated how different audio sampling rates during fine-tuning impact classification accuracy.
Key insights
- Models pre-trained with higher frequency resolution (e.g., 32 kHz) exhibit greater robustness and generalize effectively to fine-tuning data with varying, including lower, sampling rates.
- Knowledge transfer from large-scale visual datasets (ImageNet) can be highly effective for acoustic classification tasks, sometimes outperforming models pre-trained on audio-specific datasets.
- The best-performing ImageNet-pretrained model, ConvNeXtV2-tiny, demonstrated more focused and interpretable activation regions (via Grad-CAM) for correct classifications compared to audio-pretrained models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.