Privacy-Preserving Active Learning for Sustainable Aquaculture Monitoring

This article presents a novel approach to building responsible, robust, and efficient AI-powered aquaculture monitoring systems that address data scarcity, privacy concerns, and the simulation-to-reality gap.

Details

The author's journey into this specialized intersection of AI began on the edge of a salmon farm in Norway, where they faced challenges of data scarcity, privacy concerns, and the simulation-to-reality gap in building computer vision models for optimizing feeding schedules. This experience led to the formulation of a three-pillar approach: 1) Active Learning to learn efficiently from minimal expert input, 2) Privacy-Preserving Machine Learning techniques like Federated Learning and Differential Privacy to enable model training without exposing raw, proprietary farm data, and 3) Inverse Simulation Verification to anchor the data-driven AI predictions to physical reality. The author shares insights from their experimentation, such as the effectiveness of Bayesian Active Learning by Disagreement and the need for a customized approach to combining Federated Learning and Differential Privacy in non-IID data scenarios common in aquaculture.