I am a pediatric cardiologist and have cared for children with heart disease for the past three decades. In addition, I have an educational background in business and finance as well as healthcare administration and global health – I gained a Masters Degree in Public Health from UCLA and taught Global Health there after I completed the program.
“Overall, then, we will view cas [complex adaptive systems] as systems composed of interacting agents described in terms of rules. These agents adapt by changing their rules as experience accumulates. In cas, a major part of the environment of any given adaptive agent consists of other adaptive agents, so that a portion of any agent’s efforts at adaptation is spent adapting to other adaptive agents. This one feature is a major source of the complex temporal patterns that cas generate. To understand cas we must understand these ever-changing patterns.”
John Holland, Hidden Order: How Adaptation Builds Complexity
The article for this week, published in Nature Medicine, is on the relatively innovative concept of swarm learning. The authors begin by reminding readers that data collection faces practical, ethical and legal obstacles and has hindered the training of AI systems. Swarm learning can obviate the need for data transfer or monopolistic data governance while jointly train AI models. In this study, histopathology images from more than 5,000 patients with hematoxylin and eosin stained pathology slides of colorectal cancer were part of a swarm learning strategy.
There is a significant difference between federated learning and swarm learning. In federated learning, multiple AI models are trained independently and share model weights but do not share data. A central coordinator does govern the learning progress based on all the trained models. In swarm learning, the AI models are trained locally and the models are combined centrally without any central coordination rendered possible by a blockchain-based coordination amongst the centers.
The study showed that the swarm learning model outperformed the local trained models while matching the performance of merged datasets. In addition, the swarm learning model was deemed to be data efficient. This swarm learning strategy, which has some elements of complex adaptive systems in which there is no central leader, can obviate the need of data transfer as it trains distributed AI models for the images. The future of swarm learning can be promising especially with patients with rare diseases that have relatively small datasets locally.
Read the full paper here: https://www.nature.com/articles/s41591-022-01768-5