Google and the quest to bring Artificial Intelligence to the masses

I recently attended the 11^th Google I/O, a conference where 7000 developers gather near Google’s headquarters to learn about, and celebrate, everything new that Google announces across its increasing range of products and services.

The announcements continued last year’s ‘Artificial Intelligence first’ trend. The number of new applications of machine learning techniques to Google’s products could now be qualified as ‘mind-boggling’, from neural beamforming microphones that helps reduce Google Assistant’s speech recognition error rate to under 5% (it was 23% in 2013), to improving night-time photography; in fact it is reaching a point where it’s hard to find an area now where Google is not applying data science to improve the quality of experience for their users.

What I found most interesting about this transition to AI first is that Google is democratising, not just the benefits of AI for all their users, but also access for developers to much of the new supporting technology. From frameworks and APIs (TensorFlow / TensorFlow Lite and Keras), training data, to computing resources with Google Cloud and the new TPU2, developers now have a catalogue of tools for development of machine learning applications at their disposal that no one could have dreamed of just a couple of years ago. This message was repeated throughout the event but the profound implications for all of us will take time to become clear.

Take Google Lenses, one of the most remarkable announcements – in my opinion – of the event. This camera extension uses a combination of computer vision, geolocation and machine learning to provide contextual search in a broad range of fields, truly augmenting the world around us. What could be the impact of such a tool in a few years? I’m old enough to remember the time before Google when a question would pop up, for which I had no answer, and that was fine. When my son is older, he may remember the time when he saw a painting or a building, and didn’t immediately had overlaid on it the author, date it was made and every other bit of relevant information. The combination of camera sensors, computing and graphic power, connectivity and machine learning are making information accessible in ways that even today seem like magic.

Progress doesn’t stop here. Nowadays, mobile users (all of us) take revolutions like this for granted but underneath the technology sector continues to push the boundaries to make it possible. Future applications will demand even better mobile technology, in two ways:

• The progress of AI science is so staggering that current hardware can’t keep up with the demands. Inference is the process of applying a trained machine learning model to a problem. Much of the inference for mobile applications, from hand writing recognition to the application of artistic filters to your pictures, happens on the device itself (as opposed to model training, which requires vast amounts of cloud computing power). Faster and more efficient CPUs, GPUs and accelerators will be required to execute ever more, and more complex machine learning mobile applications.
• At the same time, the next billion mobile users will come from the developing world, where infrastructure, connectivity and the cost of devices are much more limited. The industry needs to create cost-effective technology that satisfies the needs of this generation of users and brings them to the information era as first-class citizens.

Who knows where the path of technological progress is taking us. Some applications like Google Lens will work a lot better in the ‘smart glasses’, mixed reality form-factor that already feels imminent. Others, will come in devices that we haven’t imagined yet. Either way, they will have to pack increasing amounts of compute power into smaller physical and power constraints, something that Arm has been doing for the past 25 years, and will continue to do in the future, with a commitment to increasing AI computational performance by up to 50x in the next three to five years. This week Arm has announced more details on how we plan to achieve this.

What is certain is these future devices will be a lot smarter than today smartest-phones thanks to the progress of AI and the hardware to power it in the cloud and in mobile. Will Google reach the point where there is no more machine learning left to sprinkle on their products? I wouldn’t bet for it any time soon.