ComnetCo Blog

A new AI model allows researchers to share insights, not data.
In the world of artificial intelligence (AI), there’s a new kid on the block. As if all the myriad branches of AI were not confusing enough, in addition to everything from deep learning to fuzzy logic we now have “swarm learning.” As a form of machine learning, it basically facilitates training models at the edge, so the edge devices get smarter and also train their peers.
But swarm learning also puts two new twists on standard machine learning that make it very exciting for a range of applications: it works as a decentralized model, and it links the edge devices with blockchain technology. This means researchers can share insights without sharing data, thus enabling collaboration while preserving privacy.
A natural evolution of ML
Actually, it is not a completely new concept but evolves from other forms of AI. The journal Nature debuted swarm learning in May 2021, and the authors described the concept as the fourth step in a progression of machine learning concepts. First, there is local learning where data and computation reside at different, disconnected locations. Next, comes cloud-based centralized learning. In the third evolution, federated learning, computing is performed at the point where data is created, collected, and stored with parameter settings orchestrated by a central parameter server.
In the fourth phase of this evolution, swarm learning makes it possible to share just the neural network inference data from many distributed edge nodes all linked over a blockchain. In other words, by sharing insights derived from AI analytics performed at the edge, researchers can collaborate in different jurisdictions without sharing the actual data. This new distributed approach eliminates the need for centralized coordination and a parameter server; a potential threat vector for bad actors to corrupt or manipulate confidential data. The individual edge nodes almost literally become a swarm, exchanging parameters for learning securely using blockchain technology.
Concept of Swarm Learning
Increase accuracy and reduce biases in AI models
In a recent ComnetCo / Hewlett Packard Enterprise (HPE) white paper, Dr. Eng Lim Goh, HPE’s Senior Vice President & Chief Technology Officer for Artificial Intelligence, described how “More and more, we are thinking at some point a smart edge device should not only be running a trained AI/ML learning model given to it by humans, but should also be doing learning on its own based on the data it’s collecting,” explained Dr. Goh. “This is the next forward-thinking concept.”
Dr. Goh also described to us his recent collaboration with the World Health Organization (WHO) on the potential for swarm learning to solve a huge challenge in medicine.
Medicine is an inherently decentralized field. Hospitals around the globe want to utilize the massive amounts data collected from edge devices within the world of medical IoT. “However, one catch here is that each sensor will be looking at its own compartmentalized data, and therefore will be highly biased towards the data it’s seeing,” explained Dr. Goh. “Eliminating this bias by sharing AI/ML outcomes from many edge devices is one reason why we came up with a concept called ‘Swarm Learning’.”
Swarm learning leverages the security of blockchain smart contracts to work collaboratively with peers and improve model insights. In fact, the authors of the original Nature article showed that swarm learning classifiers outperformed those developed at individual sites.
In addition to better accuracy, swarm learning is also more efficient. By putting machine learning at the edge, or the near edge, the data remains at the source preventing the inefficient movement of data—or data duplication—to the core or central location.
Enabling collaboration while protecting privacy
The beauty of swarm learning is that it allows for the insights generated from data to be shared without sharing the source data itself. Data is not moved from the sources thus preserving data privacy by limiting data movement.
As in federated learning, the machine learning method is applied locally at the data collection source. Only inferred insights from that data are shared between the nodes. Protecting privacy by not exposing patient private data is not only critical for maintaining compliance with data privacy laws but is also a basic duty of researchers. According to the National Institute for Health (NIH), “Protecting patients involved in research from harm and preserving their rights is essential to ethical research.”1
This new approach means, for example, that hospitals can share insights derived from applying AI at the edge without actually risking exposure of patient protected data to bad actors. Plus, there is no central custodian that aggregates all the data, leveraging a blockchain helps to ensure data integrity. This protection is not only critical for protecting patient privacy but also safeguarding confidential data from the probing noses of hackers, such as, for example, nation states spying on vaccine research. This decentralized, distributed model shares only the insights gleaned from the data, often derived using AI machine learning models.