Skip to content

Module 4: Advanced Methods and Best Practices in FAIR AI & ML

Module Summary

This module explores cutting-edge AI paradigms and essential data management practices in modern ecological research. Building on core machine learning and computer vision fundamentals, students will investigate how advanced AI methods enable species identification, trait analysis, segmentation, and synthetic image generation with limited labeled data. Alongside these computational techniques, the module emphasizes practical data stewardship throughout the project lifecycle, equipping students with hands-on experience in documentation standards to ensure all digital research outputs are Findable, Accessible, Interoperable, Reusable (FAIR), and fully reproducible.

Week 1 - Advanced Methods in AI and ML

Class Date: October 27, 2026

Instructors: Harry Chao, Cheng Zhang

Recording from 10/27 synchronous session

Lesson Summary

This lesson will introduce advanced AI methods for ecological and biological research, including vision foundation models, multimodal learning, generative AI, and agentic AI. Building on the previous introduction to ML and CV, students will explore how pretrained models can support species identification, biological trait analysis, image segmentation, and species generation with limited labeled data. We will also discuss how AI agents can integrate models, scientific datasets, and tools to support ecological research and discovery.

Learning Objectives

At the end of this lesson, students will be able to:

  • Explain foundation models and their advantages over traditional task-specific ML models.
  • Describe how vision and multimodal foundation models support species identification, segmentation, and biological trait analysis.
  • Understand how generative AI can support biodiversity research and recognize its limitations.
  • Detail how AI agents can combine models, data sources, and tools for ecological research.

Pre-Lecture Activities and Reading

due by 2pm ET on Tuesday, October 27, 2026

Watch

  1. Pre-lecture for advanced AI and computer vision for science
  2. Building scalable and grounded AI for Science

Read

Advanced ML and CV

  1. Interpretable AI
  2. Advanced AI/ML for ecology and biodiversity

Multimodal AI (vision taxonomy models)

  1. BioCLIP: A Vision Foundation Model for the Tree of Life
  2. BioCLIP 2: Emergent Properties from Scaling Hierarchical Contrastive Learning

Vision Foundation Models

  1. SAM 3: Segment Anything Model 3 original paper, SST for specimen image segmentation
  2. DINO v3: Self-supervised learning for vision at unprecedented scale (optional)

Generative AI

  1. Read Chapter 32 (generative models) Foundations of Computer Vision
  2. Stable Diffusion
  3. TaxaAdapter: Scaling Fine-grained Species Image Generation To the Tree of Life
  4. TaxaDiffusion: Progressively Trained Diffusion Model for Fine-Grained Species Generation (optional)

Interpretable AI

  1. Interpretable and Testable Vision Features via Sparse Autoencoders
  2. Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation (optional)

Agentic AI

  1. Autonomous Scientific Discovery via Iterative Meta-Reflection

Post-class Assignments

  1. Journal Entry

    • Prompt: Choose an ecological or biological research question that interests you.
      1. How might one or more of the AI approaches introduced in this lesson - foundation models, multimodal AI, generative AI, or agentic AI - help investigate this question?
      2. Describe what data you would use, which AI tools or models might be appropriate, and what information you hope to obtain. What are the potential limitations or risks, and how would you determine whether the AI-generated results are scientifically reliable?

Class Session Office Hours

Who: Jake Beattie

Date: Thursday, October 29, 2026

Time: 2pm Eastern Time

Zoom: Office Hour Zoom Link

Week 2 - Best Practices for FAIR AI and ML

Class Date: November 3, 2026

Instructors: Hilmar Lapp, Elizabeth Campolongo

Recording from 11/03 synchronous session

Lesson Summary

This lesson aims to prepare students for proper data management and documentation for ensuring all digital products are Findable, Accessible, Interoperable, Reusable (FAIR) and Reproducible. It provides hands-on experience with aspects of the project life cycle and data documentation.

Learning Objectives

At the end of this lesson, students will be able to:

  • Explain the importance of the FAIR principles for research.
  • Apply the understanding of FAIR principles to documentation practices (e.g., how data was created, formatted, annotated, keywording, version control).
  • Understand their role and responsibilities in the research project life cycle.
  • Understand what goes into a dataset card and what differentiates a helpful dataset card from an unhelpful dataset card.
  • Apply that understanding to fully document their datasets.

Pre-Lecture Activities and Reading

due by 2pm ET on Tuesday, November 3, 2026

Watch (required):

  1. Demystifying Research: Reproducibility Made Realistic Seminar on Oct. 27 at 12pmET (session will be recorded, but still requires registration): Register

Read

  1. Imageomics & ABC FAIR Guide
  2. Project Lifecycle Diagram
  3. Balk, M. A., Bradley, J., Maruf, M., Altintaş, B., Bakiş, Y., Bart, H. L. Jr, Breen, D., Florian, C. R., Greenberg, J., Karpatne, A., Karnani, K., Mabee, P., Pepper, J., Jebbia, D., Tabarin, T., Wang, X., & Lapp, H. (2024). A FAIR and modular image-based workflow for knowledge discovery in the emerging field of imageomics. Methods in Ecology and Evolution, 15, 1129–1145. https://doi.org/10.1111/2041-210X.14327.
  4. Shehu A, Nussinov R. Four questions for AI-ready biological data. Nature Machine Intelligence, 8,1017–1019 (2026). https://doi.org/10.1038/s42256-026-01270-4.
  5. Kline J, Meier K, Shukla V, Rolland EGA, Iannino E, Laporte-Devylder L, Catricheo CAM, Costelloe B, Campolongo E, Midtiby HS, Tuia D, Risse B, Lundquist UPS, Christensen AL, Remondino F, Richardson T, Berger-Wolf T (2026) FAIR^2 drones: An AI-ready standard for cross-domain wildlife drone datasets. arXiv [cs.RO], https://doi.org/10.48550/arXiv.2606.00355.
  6. White M, Haddad I, Osborne C, Yanglet X-YL, Abdelmonsef A, Varghese S (2024) The Model Openness Framework: Promoting completeness and openness for reproducibility, transparency, and usability in artificial intelligence. arXiv [cs.LG], http://arxiv.org/abs/2403.13784.
  7. Sohn E (2023) The reproducibility issues that haunt health-care AI. Nature, 613(7943):402–403. https://doi.org/10.1038/d41586-023-00023-2.

Complete

  1. Data Training (View complete training resources and instructions, video)
  2. This is a group assignment (Assignment link) - working with your PALS group members1, create a dataset card for a proposed dataset:
    • Download a manageable dataset (of any size and about any subject) from the NEON Data Portal.
    • State explicitly what the data represent; and
    • Work as a group to construct a dataset card (using the dataset card template); dataset cards created by your PALS group will be reviewed in class.
    • Use NEON Tutorials as needed

Post-class Assignments

  1. Journal Entry
    • Prompt: Please reflect on your experience going through the Dataset Card generation and review process. Reflecting on the Digital Product Policy and Life Cycle, which steps or aspects do you find least intuitive/most challenging/least obvious to be important? Where did you think you would spend the most time vs where you actually spent the most time? What new approaches or strategies will you incorporate into your future projects?

Class Session Office Hours

Who: Net Zhang

Date: November 5

Time: 2pm ET

Zoom: Office Hour Zoom Link

  1. Groups of 3 have also been assigned for those not signed up for PALS. ↩