Federated Learning and Its Applications

Key takeaways

This project was an immersive exploration of deep learning, computer vision, federated learning, and machine learning. I implemented and evaluated a federated learning pipeline using a U-Net model for wheat head instance segmentation. I trained models on ARC remote servers, worked on image processing and transformation techniques, and ran a literature review that sharpened my research toolkit.

Paper

Download the draft manuscript: FL draft manuscript

Abstract

In my final year thesis project at the University of Calgary, we tackled the challenges associated with centralized learning in wheat head instance segmentation, specifically addressing privacy concerns and computational limitations. The goal was to improve accuracy and efficiency in this agricultural task by leveraging federated learning, an approach that improves data privacy and avoids centralized bottlenecks.

Federated learning solution

Federated learning emerged as a solution to these challenges, and our project focused on implementing and evaluating a federated learning pipeline for wheat head instance segmentation. The pipeline was model-agnostic and used a modified U-Net to compare approaches across clients.

Comparison of approaches

We compared two approaches - FedAvg and FedProx - and evaluated using Intersection over Union and Dice scores, against a centralized baseline. Wheat head images from 18 institutions were sourced from the Global Wheat Head Dataset. FedAvg achieved a Dice score of 0.9355 on the test set, outperforming the centralized baseline.

Conclusion

This project shows the potential of federated learning to address privacy and scalability challenges in agricultural computer vision. It is a practical path forward for applying ML where data access is constrained.

Person at work … check back later.