Artificial intelligence is relentlessly developing; it keeps moving fast. The modern era of deep Machine Learning began only eight years ago, at the 2012 ImageNet Competition. Since then, new methods have been proposed, but the rapid pace of innovation turns current solutions into outdated ones within a few years. Researchers have begun to rely on computers to make “intelligent” decisions in the way humans do.
Machine learning refers to one aspect, which is designing algorithms that “learn” to generalize from past experiences to perform better on future tasks.
However, according to experts, the future of AI belongs to unsupervised learning.
Supervised machine learning has represented a remarkable driver of AI progress. Methods belonging to supervised learning provide the algorithm with value that needs to be output for each observation. One of the main limitations of this process is that algorithms focus only on categories that researchers have identified a-priori.
On the other hand, unsupervised learning is an approach in which algorithms themselves learn from data without human guidance – experts do not define the classes before. In a nutshell, the system learns from the training data set about some characteristics by observing the behavior of relationships between entities. Unsupervised learning emulates the way human-beings learn about the world.
Among the challengers of the digital world, we find data privacy. Data is considered the ‘new oil’, the lifeblood of AI. Nonetheless, privacy issues play a limiting role. In 2017, researchers at Google contrived the concept of ‘federated learning’. According to Forbes, over the past year, interest in federated learning has become prominent. Today, the standard approach to building ML classifiers is to gather all the training data in one place and then to train the model on the data. Inevitably, this approach is not feasible as a large amount of the world’s data cannot be moved to a single central data repository for privacy and security reasons. Federated learning converts such conventional approach to AI.
In federated learning, data is left where it belongs to – devices and servers – and is trained locally. The classifier is instead delivered to each device. In the end, the final outcome is the aggregation of the resulting parameters from each device. The upshot owns the same efficiency as if data had been trained is one single cloud, but actually they had not.
A practical usage of this approach is in the health care systems. Indeed, patients’ personal information is of utmost sensitivity. AI federated learning tools enable experts to develop attainable health records, avoiding privacy breaches.