Many consumer tech companies rely on artificial intelligence and machine learning.

Federated learning is now used, which is a form of machine learning that uses algorithms installed on devices that are part of a network: in this way the data does not leave the device.

Given the particular attention paid to privacy, this type of learning could become a very popular and preferred method.

The decentralization of learning

The standard used until now provides that the data on tablets, smartphones, laptops etc. etc. were delivered to a centralized server and from there the algorithm drew its information.

Following research, the technicians found that the server, where the learning algorithm is inserted, does not need to be in a loop.

The exact opposite is possible: each device that collects data is equipped with the learning algorithm. In this case, the sensitive data, which could be the location, does not leave the device, but is processed immediately by the algorithm.

This is very similar to what happens for the validation of the blockchain.

In the federated learning system, there are two management options: one-party and multi-party management.

The difference is basically on the management of the devices that are part of the network.

One-party management means that the devices belong to a single company, while multi-party management consists of a partnership of companies.

How federated learning works

First of all, the algorithm must be studied on a central server so that the learning operations are correct. Once Machine learning has been defined, the algorithm is ready to be distributed to the network nodes and activated.

The procedure consists of 5 steps.

1 - Sending the algorithm to the network nodes;

2- Each archived model begins learning by analyzing the data generated by the device;

3 - The learning results of the algorithms are sent from the devices to the central server;

4 - The server learns the various results and recodes the algorithm more precisely thanks to the various types of algorithm that have been received;

5 - Redistribution of the new algorithm to all devices

Let’s have a look to tangible examples of federated learning, for example Netflix: at each access it proposes new titles according to the genres that the user has seen.

Not only that, other variables (which do not reach the central server) contribute to the construction of the list, such as age, gender and position.

Advantages

As we have seen, all the data that are used for learning, do not abandon the device that generated them, but only the results are communicated in order to refine the algorithm that will subsequently be updated.

To this important aspect of privacy, it is also added that:

- Reduction of costs for sending data sets between the device and the server including the reverse path for receiving the processed data;

- Network latencies are low and in some cases even non-existent.

As we have seen, the decentralization of AI has extremely positive sides as the network of devices that contribute to the implementation of the algorithm are destined to increase, while maintaining the exchange of data to an acceptable extent.

Secondly, a certain privacy of personal data is maintained.