AI for Biodiversity and Ecosystems

The theme for day three is AI for biodiversity and ecosystems. Partners such as Google DeepMind, GitHub, CIFAR, and Mila have supported the summer school. Information about the organization and community platform is provided. Housekeeping checks are explained for asking questions.

Guidelines for asking questions during the lecture are detailed. Information about the two lectures for the day on AI for wildlife conservation is given.

Professor Davis Tuya's background and expertise in wildlife conservation are highlighted. The introduction and thanks for his participation are expressed.

Professor Davis Tuya's educational background and work experience are shared, including his leadership role at EPFL.

Professor Tuya expresses gratitude for the opportunity to speak and shares his excitement. He briefly discusses the topic of biodiversity conservation and the importance of collaboration.

An interactive poll for the audience to share their backgrounds and interests is conducted, focusing on ecology, biodiversity, and AI.

The speaker discusses the importance of biodiversity conservation, highlighting the impact of species extinction and the services provided by ecosystems.

The global issue of habitat loss is addressed, emphasizing the widespread impact on ecosystems worldwide.

The challenges in species data collection and conservation efforts are outlined, emphasizing the need to address data-deficient species for effective conservation.

The creation of a global biodiversity framework and its key goals to conserve protected areas and reduce biodiversity decline are discussed.

The speaker discusses the role of technology, including mobile sensors, camera traps, drones, and satellite data, in wildlife conservation and monitoring. Various sensors and their applications are explained.

The application of machine learning in wildlife monitoring and species detection is explained. The process of training AI models to detect and analyze animal populations is detailed.

Discussion on deep learning models for semantic segmentation, counting, and classification.

Explanation of Mega detector software developed by colleagues at MITsariri for object detection in images from camera traps.

Examples of using Mega detector in wildlife protection solutions and Department of fishing game for animal and vehicle detection in images.

Explanation of a method to speed up animal detection in wildlife census using AI models and human input.

Description of an interactive process involving a model developed by Alexandria Del Blanc for animal detection and retraining.

Discussion on the comparison between human and AI model performance in animal detection with insights on missed detections.

Optimization methods are used in various applications such as finding the shortest path on Google Maps and scheduling airline flights. In conservation, optimization tools like Marxan software help pick relevant lands for purchase based on budget constraints.

Machine learning algorithms, based on optimization principles, are used for tasks like classifying data points accurately. In conservation, optimization aids in decision-making for tasks such as selecting specific parcels of land for habitat conservation.

The predict and optimize paradigm combines data prediction with optimization tools. In conservation, this approach is used for tasks like predicting poacher behavior and optimizing ranger patrols based on remote sensing data.

Market design involves matching systems like rideshare platforms and kidney donor assignments. In conservation, it is applied in payments for ecosystem services to compensate landowners for land conservation efforts.

Conservation faces challenges in allocating payments for ecosystem services effectively due to distinct players with different incentives. Addressing positive externalities and designing auctions for land purchase are key considerations in conservation market design.

Online learning is utilized in tasks like email spam filtering and ranger patrol planning in protected areas. It involves learning incrementally from new data samples and making decisions based on real-time information.

Reinforcement learning operates in dynamic settings, learning from actions and rewards in real time. It has applications in conservation for tasks like fisheries management and precision agriculture, optimizing decisions to impact long-term outcomes.

Causal inference is crucial in understanding cause-and-effect relationships, especially in conservation interventions like payments for ecosystem services. It helps differentiate correlation from causation in conservation decision-making.

Managing uncertainty is essential in conservation decision-making to quantify uncertainty in species distribution models. Strategies like robust optimization help incorporate uncertainty into decision-making processes effectively.