If there’s one resource the world isn’t going to run out of anytime soon it’s data. International analyst firm IDC estimates the ‘Global Datasphere’ – or the total amount of data stored on computers across the world – will grow from 33 zettabytes in 2018 to 175 zettabytes in 2025. Or to put that in a more relatable form, 175 billion of those terabyte hard disks you might find inside one of today’s PCs.
That data pool is an enormous resource, but one that’s far too big for humans to exploit. Instead, we’re going to need to rely on deep learning to make sense of all that data and discover links we don’t know even exist yet. The applications of deep learning are, according to Intel’s AI Technical Solution Specialist, Walter Riviera, “limitless”.
“The coolest application for deep learning is yet to be invented,” he says.
So, what is deep learning and why is it so powerful?
Teaching the brain
Deep learning is a subset of machine learning and artificial intelligence. It is specifically concerned with neural networks – computer systems that are designed to mimic the behaviour of the human brain.
In the same way that our brains make decisions based on multiple sources of ‘data’ – i.e. sight, touch, memory – deep learning also relies on multiple layers of data. A neural network is comprised of layers of “virtual or digital neurons,” says Riviera. “The more layers you have, the deeper you go, the cleverer the algorithm.”
There are two key steps in deep learning: training and inference. The first is teaching that virtual brain to do something, the second is deploying that brain to do what it’s supposed to do. Riviera says the process is akin to playing a guitar. When you pick up a guitar, you normally have to tune the strings. So you play a chord and see if it matches the sound of the chord you know to be correct. “Unconsciously, you match the emitted sound with the expected one,” he says. “Somehow you’re measuring the error – the difference between the two.”
If the two chords don’t match, you twiddle the tuning pegs and strum the chord again, repeating the process until the sound from the guitar matches the one in your head. “It’s an iterative process and after a while you can basically drop the guitar, because that’s ready to go,” says Riviera. “What song can you play? Whatever, because it’s good to go.”
In other words, once you’ve trained a neural network to work out what’s right and wrong, it can be used to solve problems that it doesn’t already know the answer to. “In the training phase of a neural network, we provide data with the right answer… because we know what is the expected sound. We allow the neural network to play with that data until we are happy with the expected answer,” says Riviera.
“Once we’re ready to go, because we think the guitar is playing well, so the neural network is actually giving the expected answer or the error is very close to zero, it’s time to take that brain and put it in a camera, or to take decisions in a bank system to tell us that it’s a fraud behaviour.”
Deep learning as a concept isn’t new – indeed, the idea has been around for 40 years. What makes it so exciting now is that we finally have all the pieces in place to unlock its potential.
“We had the theory and the research papers, we had all the concepts, but we were missing two important components, which were the data to learn from and the compute power,” says Riviera. “Today, we have all of these three components – theory, data and infrastructures – but what we’re missing is the fourth pillar, which is creativity. We still don’t know what we can and can’t achieve with deep learning.”
That’s not to say that deep learning isn’t already being put to amazingly good use.
Any regular commuter will know the sheer fist-thumping frustration of delays and cancelled trains. However, Intel technology is being used to power Resonate’s Luminate platform, which helps one British train company better manage more than 2,000 journeys per day.
Small, Intel-powered gateways are placed on the trackside, monitoring the movements of trains across the network. That is married with other critical data, such as timetables, temporary speed restrictions and logs of any faults across the network. By combining all this data and learning from past behaviour, Luminate can forecast where problems might occur on the network and allow managers to simulate revised schedules without disrupting live rail passengers. The system can also make automatic adjustments to short-term schedules, moving trains to where they are most needed.
The results have been startling. On-time arrivals have increased by 9% since the adoption of the system, with 92% of trains now running to schedule.
Perhaps just as annoying as delayed trains is arriving at the supermarket to find the product you went there for is out of stock. Once again, Intel’s deep learning technology is being used to avert this costly situation for supermarkets.
The Intel-powered Vispera ShelfSight system has cameras mounted in stores, keeping an eye on the supermarket shelves. Deep-learning algorithms are used to train the system to identify individual products and to spot empty spaces on the shelves, or even products accidentally placed in the wrong areas by staff.
Staff are alerted to shortages using mobile devices, so that shelves can be quickly restocked and lost sales are kept to a minimum. And because all that data is fed back to the cloud, sales models can be adjusted and the chances of future shortages of in-demand products are reduced.
Only the start
Yet, as Riviera said earlier, these applications of deep learning are really only the start. He relays the story of the Italian start-up that is using deep learning to create a system where drones carry human organs from hospital to hospital, eliminating the huge disadvantages of helicopters (too costly) and ambulances (too slow) when it comes to life-critical transplants.
It’s not the only life-saving application he can see for the technology, either. “I’d like to see deep learning building an autonomous system – robots – that can go and collect plastic from the oceans,” he says. “We do have that capability, it’s just about enabling it and developing it.”
“The best [use for deep learning] is yet to be invented,” he concludes.