“Machine learning is a core, transformative way by which we're rethinking everything we're doing.” Interesting words and current for our time, made even more important by the man who uttered it - Sundar Pichai, CEO of Google.

At a recent machine learning event held at Google HQ in Bryanston, Blaise Agüera y Arcas, principal scientist for Machine Learning and AI at Google, was quick to point that what Pichai said is literally true. Machine learning forms part of every Google product, and Agüera y Arcas and his team of 200-300 people work closely with almost all divisions within the company.

Neural nets are key

The development and deployment of neural networks is one part of Agüera y Arcas domain, with these networks powering machine learning. This in turn drives, for example, image recognition of users’ photos in Google Photos, natural language processing in the Google Assistant, or machine translation in Google Translate.

But while one would think that neural networks would reside exclusively in large data centres, Agüera y Arcas notes that Google is also working on neural nets that will run locally on users’ smartphones. It comes across as counter intuitive that neural networks should move away from the cloud, but he believes there are benefits to having these running on-device, including better privacy, response time and the fact that you don’t need a stable connection to the internet for it to work.

“My own belief is that ultimately we’re going to have little sorts of companions that should be our own and not the extension of Google or some other company,” he said. Indeed, Laura Scott, Reputation Communications and Public Affairs at Google, later emphasised the utmost importance of the Google Assistant and the company’s quest to build the ultimate virtual helper. One that should understand not only the world (who is Nelson Mandela?), but your world (what restaurants are there at Nelson Mandela Square?), and your current context (how do I get there?).

Getting back to the on-device neural networks that Agüera y Arcas and his team are working on, some of these can run on current smartphones and are large and complicated enough to compete with the visual cortex of a mouse or a rat. One project his team is currently busy with, is a locally residing bird identification platform, that can not only locate the bird in a picture but correctly identify the species too.

The training needed

If these neural networks run locally and not in the cloud, where then would the big data come from to improve and further train these? Agüera y Arcas points to a technique they developed called federated learning, with learning also being distributed amongst devices.

If you do something on your device, say select text, this is one small bit of data in a broader set of ‘how people select text in Android’, and can be used to make everyone’s experience better. The text that you selected - did you have to adjust it to include a complete address, the dialling code in front of the telephone number, or the whole restaurant name?

Your device remembers this fix and makes local changes, but from time to time that data is compressed, encrypted and uploaded to the cloud. This is combined with everyone else’s changes and sent back down to each smart device to improve the neural network of all devices. Agüera y Arcas believes this allows everyone’s selections to always remain private, and the learning that happens to be almost bigger than big data because every alteration made on a local device forms part of the broader training on all devices.

No, A.I. won’t wake up

Looking at the broader existential implications of Artificial Intelligence (A.I.), Agüera y Arcas does not quite buy into the belief of A.I. becoming sentient, calling it unfounded. “I don’t know whether we will ever be able to make a machine that can feel, a machine that is conscious or a machine that can even fully reason or have common sense”. He does, however, state that there are threats that AI pose, including the rise of economic inequality, bias in the construction of AI and issues around labour and privacy.

Having said that, the changes that are brought about by machine learning in some of Google’s products already are simply astounding. Try for example the Google Translate app that translates words in real-time via augmented reality, and you might just find that this sufficiently advanced technology is indeed indistinguishable from magic. And with daily advances being made in machine learning, Agüera y Arcas and his team of modern day alchemists are set to bring even more magic to the Philosopher’s Stone you call your smartphone.

Check Agüera y Arcas’  TED Talk on on machine creativity below.