Clustering the BioTrove Dataset

Clustering the BioTrove Dataset was featured in the 2025 Machine Learning Marathon (MLM25). This challenge asks participants to discover genus- and species-level structure in biodiversity images using unsupervised and self-supervised learning methods.

Challenge design

• Task: Cluster biodiversity images to recover taxonomic structure (genus and species groupings) without explicit labels.
• Domain: Biodiversity and ecology – automated species identification can support pest control, crop monitoring, biodiversity assessment, and environmental conservation.
• Data: Images drawn from BioTrove, the largest publicly accessible biodiversity image dataset (161.9 million images, ~366K species), curated from iNaturalist with research-grade annotations.
• Methods: Contrastive learning, autoencoders, CLIP-based embeddings, and other unsupervised/semi-supervised approaches.

Links

• Kaggle challenge: Clustering the BioTrove Dataset
• Winning writeup: 1st place: It All Depends on a Good Embedding
• BioTrove project: baskargroup.github.io/BioTrove

Related resources

• Data: BioTrove: Learn more about the BioTrove dataset, including its construction from iNaturalist and available CLIP embeddings.
• Data: iNaturalist: The citizen-science platform underlying BioTrove’s research-grade biodiversity observations.
• Library: Hugging Face: BioTrove-CLIP models and the dataset are hosted on Hugging Face.
• Workshop: Intro to Deep Learning with PyTorch: Build foundational deep learning skills relevant to contrastive learning and embedding-based clustering.
• Paper: BioTrove: A Large Curated Image Dataset Enabling AI for Biodiversity (NeurIPS 2024)

Questions

If you have any lingering questions about this project, please feel free to post to the Nexus Q&A on GitHub. We will improve materials on this website as additional questions come in.