Reproduce “BIOSCAN-5M: A Multimodal Dataset for Insect Biodiversity” 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
Zahra Gharaee, Scott C. Lowe, ZeMing Gong, Pablo Millan Arias, Nicholas Pellegrino, Austin T. Wang, Joakim Bruslund Haurum, Iuliia Zarubiieva and 5 more
Researchers introduce BIOSCAN-5M, a multimodal dataset comprising over 5 million arthropod specimens, each with high-resolution images, DNA barcodes, and hierarchical taxonomic labels. The dataset supports advanced machine learning for biodiversity, enabling DNA-based classification models to achieve 99.28% species-level accuracy and improving multimodal retrieval tasks.
Key insights
- Large-scale pretraining on DNA barcodes from BIOSCAN-5M significantly improves both closed-world species-level classification and open-world generalization to unseen genera.
- DNA barcodes are highly effective for fine-grained taxonomic clustering, with DNA embeddings demonstrating substantially higher agreement (up to 91% AMI at species level) with ground-truth labels compared to image embeddings.
- Contrastive learning successfully aligns image, DNA, and textual taxonomic embeddings into a shared space, facilitating improved cross-modal retrieval and demonstrating the continued benefit of dataset scale for multimodal tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.