What is it about?
Today’s artificial intelligence (AI) infrastructures often depend on centralized servers, raising concerns about privacy, fairness, and reliance on a few large providers. Decentralized machine learning offers an alternative by allowing devices to collaborate directly, but it struggles with balancing personal accuracy and general performance. Our work tackles this challenge in two ways: first, we use blockchain to provide trust and accountability among devices without relying on a single authority; second, we design a low-cost and test-time adaptation method that lets devices keep using their personalized model during testing, but automatically switch to a shared, more general model whenever they encounter unfamiliar cases. This makes decentralized AI both practical and scalable—paving the way for more secure, adaptable, and user-centric intelligence at the network edge.
Featured Image
Photo by Shubham Dhage on Unsplash
Why is it important?
Most AI today relies on large cloud servers, which raises concerns about privacy and fairness. Decentralized machine learning enables devices like phones to train models locally, but it often struggles to strike a balance between personal accuracy and general reliability. Our approach uses blockchain to build trust among devices and introduces a fast test-time adaptation method, where each device uses its own model but switches to a shared model only when it faces something unfamiliar. This makes AI more private, efficient, and adaptable for real-world use.
Perspectives
What excites me most about this work is finding a way to make powerful AI models more practical and accessible in everyday settings. Retraining models from scratch is costly and time-consuming, so I was motivated to explore test-time adaptation as a lightweight alternative. For me, the real value lies in showing how decentralized learning, combined with blockchain, can help strike a balance between individual needs and collective reliability. I hope this article encourages others to think about how we can make AI not only more efficient and trustworthy, but also more resilient to the challenges of real-world data shifts.
Yao Du
The University of British Columbia
Read the Original
This page is a summary of: Decentralized Model Selection for Test-Time Adaptation in Heterogeneous Connected Systems, ACM Transactions on the Web, February 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3764936.
You can read the full text:
Contributors
The following have contributed to this page
