Federated learning is a budding machine learning technique that promises powerful insights across the evolving digital frontier.
What is machine learning?
Machine learning is part of the reason your Instagram feed is flooded with products and services you didn’t know you needed, but may be oddly well suited to your interests and style. It’s why your Netflix feed populates with shows you love, or shows very similar to the ones you love. Machine learning algorithms use concepts in statistics to discern patterns from exorbitant amounts of data—data they feed on daily through our clicks, likes, saves, and even spoken input.
Creepy or clutch, when it comes to the impact of machine learning, serving up relevant content doesn’t quite hold a candle to how it can improve other sectors, say, healthcare.
When you consider one of the primary ways doctors and healthcare professionals evolve their understanding of the human body—by analyzing cases, research, and trends over time, you might assume the proverbial “they” already employ all of that existing information floating out in the ether, so anything they can figure out, they have. But that isn’t necessarily the case. While thousands of hospitals collectively aggregate millions of data points each day, until recently that information has been scattered, protected, and vastly inaccessible to those who’s analysis could be most instrumental for breakthrough treatment.
An unfortunate truth, because the technology to process that data and make meaningful predictions, for the most part, already exists. Properly employed, machine learning can utilize the ever-amassing pool of recorded health information to observe medical patterns at a scale previously unimaginable.
To be effective, machine learning algorithms need to “train” with as many (and as diverse of) data sets as possible. Classical machine learning approaches involve centralizing the training data—collecting it on a single server and feeding it into a model also on that server. It keeps everything in one place, limiting the input to what can be recorded at that source point. Why? For one, medical data is widely distributed across many hospitals, and often stored in different formats at those locations. Two, health information is highly personal, and so, necessarily, tightly protected by the law. Moving data around just isn’t secure.
Why can’t we just anonymize the data and feed it into the models freely?
Why go through all the trouble of creating these complex routes for health information when the data could theoretically be released without the patient’s name? Seems reasonable, but according to this article in 'Digital Medicine', the consensus is that genomic data and medical images are not assuredly protected this way. The paper also details how sensitive details within that information can too easily lead to patient re-identification and exposure.
As is frequently the case in human history, the obstacles presented to machine learning have actually paved the way for promising developments. Cue federated learning.
Related: A Breakdown of Our Rapid Prototyping Model
What is Federated Learning?
Federated learning, sometimes called collaborative learning, has made it possible for data to remain protected at its location while anonymously contributing to algorithms. These empowered algorithms can then affect change on a larger and more impactful scale by illuminating trends and patterns lost and hidden in smaller or less diverse groups. Federated learning does this by leaving the data local—moving the model rather than the data. This can even be performed by mobile phones; the mobile device downloads the training model, uses the data on the phone to improve it, summarizes those changes in a small update—the only piece of encrypted information sent back to the cloud. This approach allows many diverse data sets to contribute to a robust machine learning model. The collaborative training of models that feed into one is the making of a consensus model. Research on the efficacy of consensus models reveals that models trained by federated learning perform competitively with, and even better than some centrally hosted data sets.
Related: Top Databases of 2021: Everything You Need to Know
CHECK OUT OUR NEXT-GEN EDTECH SOLUTIONS
Find out how we can help your institution implement technology and meet its IT goals professionally and effectively.
In healthcare, this kind of machine learning is applied to detect tumors, enable precision medicine at large scale, and spot rare diseases—with a whole realm of new use cases taking shape.
The real beauty of federated learning is that it gives the ability of each contributor to define its own processes and privacy policies, control its own data access (or easily revoke it). It opens doors to new research on rare diseases, wherein cases were previously too few at small individual institutions. As data never has to be duplicated, the associated storage impact is lessened. It’s a somewhat complicated process that involves many components, but it’s looking like the method is also less carbon intensive than prior practices.
Federated learning involves a new landscape of parameters defining the scope and technologies used, which is no small undertaking. It will be an exciting development to watch unfold, and certainly an inspiring testament to humanity and the command of artificial intelligence.
Related: What is Rapid Application Development
Get a Free Copy of "The Innovation Cookbook"