This intensive 1.5-day summer school introduces participants to computer vision applications and common pitfalls of machine learning in animal science and to, combining lectures, hands-on practical sessions. Participants will learn how to annotate data, train models, deploy them for inference, and critically assess common pitfalls in applying machine learning to real-world animal science problems. A CPU laptop is sufficient, laptop with GPU or remote access to GPU’s is preferred.
Day 1
2 July 2026 – 09:00 – 18:00
Computer vision applications in animal science
Block 1 – Data & Annotation (2.5h)
Lecture: Participants are introduced to core computer vision tasks relevant to animal science, including image classification, object detection, instance segmentation, oriented bounding boxes (OBB), and pose estimation. Key topics include dataset design, annotation strategies, train/validation/test splits, and evaluation metrics.
Practical: Hands-on experience labelling datasets using tools such as CVAT. Tasks include image classification, object detection, segmentation, OBB, and pose estimation annotations, emphasizing best practices for data quality and consistency.
Block 2 – Model Training & Evaluation (2.5h)
Lecture: Covers training pipelines, transfer learning, data augmentation, and strategies to avoid overfitting. An overview of models is provided, including ResNet for classification, YOLO for detection/segmentation/OBB, and ViTPose for pose estimation.
Practical: Participants train models in Google Colab for classification (ResNet), detection (Yolo), segmentation (Yolo), and pose (Vitpose) estimation. Sessions include dataset balancing, evaluation of model metrics, and running inference on images and videos to assess performance.
Block 3 – Inference, Deployment & Scaling (2.5h)
Lecture: Focuses on inference pipelines, deployment environments (PC/NVIDIA, edge devices, HPC), and the trade-offs between speed, accuracy, and computational resources.
Practical: Participants build Streamlit demos and run models on live or recorded video. Conceptual pipelines for edge computing and HPC are explored for long-term or real-time analysis of animal datasets.
Day 2
3 July 2026 – 09:00 – 13:00
Pitfalls of Machine Learning in Animal Science
Block 4 – Pitfalls of Machine Learning in Animal Science (4h)
Workshop: This 4-hour hands-on workshop is designed for PhD students and post-docs applying machine learning in animal science, and for industry developers building commercial decision support tools. The format alternates short lectures for context with step by step coding notebooks, where participants run and modify code in prepared Jupyter style notebooks (Python and/or R). Using a real annotated sensor stream dataset, we will first train baseline models that intentionally ignore key safeguards, then diagnose what goes wrong, and finally implement fixes and re-evaluate performance.
Pitfalls covered include data leakage, inappropriate cross validation for repeated measurements per animal, overfitting and underfitting, class imbalance, bias, multicollinearity, and time series specific issues such as temporal dependence and split strategy. Participants will learn practical diagnostics (sanity checks, plots, and metric choice), and how to improve robustness without inflating model performance.
Bring your own laptop. Two instructors. Notebook delivered via Google Colab, requiring a browser and a Google account for access to shared notebooks and materials.