Why we want machine learning on smartphones and not in the cloud

ML Kit, Google’s machine learning SDK, now works entirely on the device in question: a real paradigm shift while much of machine learning is still dependent on cloud computing.

Machine learning is a discipline which has experienced a meteoric boom in recent years, to the point that it is used today everywhere, both at the level of professionals and consumers, and that it contributes to the functioning of many services that one would not suspect at first. But to offer certain complex services, which require significant computing power to link large sets of data in the most demanding concrete cases, it sometimes happens that non-specialized devices are no longer sufficient: machine learning then requires access to a cloud service that has much more powerful hardware than the average smartphone, even if these make giant leaps each year. This is accompanied by a number of disadvantages in terms of efficiency, with inevitable latency, but also and above all in terms of protection of personal data, with potential leaks that can be envisaged for both businesses and individuals.

According to SlashGear, this is why Google wants to transition from the cloud to an independent and autonomous system, which would run entirely locally on a smartphone. In any case, this is the direction taken by the Mountain View company with its ML Kit (for “Machine Learning Kit”), already available for two years but hitherto closely linked to the Firebase cloud platform. Google is not going to bury this duo, at least not at first. Instead, it is a standalone SDK (Software Development Kit, a set of tools for developers of a given platform, in this case ML Kit) that is launched for Android and iOS.

Autonomous, nomadic and tailor-made machine learning

Admittedly, that means it’s the user’s device that is going to have to do the heavy lifting, rather than a specially designed cloud server. But it also has advantages, including reduced latency and lack of communication with a server, which always opens the door to data leakage. A real argument for trusting applications based on machine learning, which regularly generate debate. In absolute terms, this should not really present any major drawbacks: today, the services that make use of the ML Kit (text recognition, object detection, translation, etc.) are relatively greedy in terms of resources compared to l use that professionals can make of it. This means that in theory, any current smartphone will be able to take on these tasks, even if it is not from the last two generations. Google has in any case created a codelab where it presents a translation tool similar to that already offered by Google, but which works entirely locally even on weak devices. It remains to be seen what use the developers will make of it and whether they will decide to abandon Firebase in favor of this local method, which seems quite promising for this kind of consumer use.

Leave a Reply

Your email address will not be published. Required fields are marked *