Intelligent Machines and the End of AI Winters
The Future of Artificial Intelligence as Seen from Today
Artificial intelligence has come a long way, but there is far more to come. Lucy Ingham speaks to Alex McMullan, chief technical officer of EMEA at Pure Storage, about the future development of AI
The road to the current state of artificial intelligence has been long, with significant peaks and troughs along the way.
Alex McMullan, CTO of EMEA, Pure Storage
“I think the development of artificial intelligence has been fascinating since the fifties; that first conference in Sweden,” says Alex McMullan, CTO of EMEA at Pure Storage, a leading provider of AI infrastructure. “We debate whether it's two or three AI winters since that point where we've gone from 'it's going to change the world' to 'oh, what a disaster'.”
The peaks in between, he says, were driven by technological advances, which may provide good reason to suspect another winter is ahead. But McMullan doesn’t think so. As a veteran of the technology industry, he has seen many innovations become a reality, and he has high hopes for artificial intelligence’s future development.
“The reason why I genuinely think it's different this time around is because we've got that balance of hardware and software capability,” he says. “I don't just mean Pure as a storage vendor, but also if you think about what Nvidia has done with GPUs; what Google has done with CPUs.
Creating value through collaboration in the AI industry
The other reason that McMullan believes that AI’s future is so bright is the unprecedented level of collaboration within the AI industry.
“We can debate the various ethics on some of the stories that are floating around, but when we talk to some of our private customers in the space industry, in healthcare, in travel and, yes, in foreign investment and in investment banking, they're all saying the same thing: they're actually collaborating with their competitors on the model, but not on the data, which is something you never hear in any of these industries,” he says.
“The open-source nature of most of the codebase is a much more natural way of building the accelerated global village, if you will, on that side of things, so everybody gets to benefit from the shared wisdom.
“The big learning point we've had is it's not the models or the infrastructure that make you successful, it's the data and the way you choose to use the data that ultimately lets you become more competitive.”
Getting to human-level intelligent machines in a decade
This combination of factors leads McMullan to believe that we will achieve machines with an intelligence at a human level in ten years.
“If you asked me in 2010, I'd have said not in my lifetime. But I think the rate of change within the last five years with all that hardware and software development and the rise of the collaborative side of that theme, it's become a much more realistic prospect,” he says.
“If you look at singularity theory, I think genuinely we're already at reasonably good narrow AI. I don't just mean the autonomous driving cars – I mean we've been doing aeroplane autopilots for quite a long time in a very closed system, but it's the branch between narrow and general AI is basically how the machine responds to unexpected events that it didn't get programmed for.
“If you asked me in 2010, I'd have said not in my lifetime.”
“A large part of that is memory. And as we have bigger and bigger storage systems, and higher and higher-performing networks with that better software, machines learn more by remembering more, and that's the natural tipping point.”
However, this does not mean he thinks humanoid robots will be amongst us.
“I don't think we're ever going to get to that Mr Data or Terminator 2 kind of thing with a man,” he says. “An intelligent machine is not going to be like us, only smarter. It's just going to be a massively capable partner. And I think that's a much better utopian vision as well, frankly.”
Healthcare: the next big winner in the AI revolution
When it comes to particular industries that are set to benefit from advances in AI, there are many, from financial services to travel, that will be enriched by the technology. But for McMullan, the industry with the most potential is healthcare and medicine.
“If you have a long-term medical record, in many cases that shows up in a shopping trolley, and as a doctor, a diagnostician, how are you supposed to assimilate that data, summarise it when it's like your Saturday shop?” he says.
“That just doesn't make sense. So there are things we can do that are really easy in that space, but for me I think the diagnosis, the predictive side of things, understanding really everything from epidemiology through to oncology, how that develops over time, is a huge benefit.
“Those, I think, are the most straightforward, easy problems to solve in healthcare. That's where I spend a lot of my time personally, because I would rather do that than work with financial services because I think that's much more interesting.”
“It's one of genuinely the easiest wins that we can actually make.”
This healthcare revolution isn’t far away – in fact, according to McMullan it is far closer than many may think.
“I would say two to three years is a realistic timeframe. We already work closely with a lot of the medical software companies, particularly in the US: all of the Epic [Systems] folks who do all of the US private hospital record management, we have strong links with them already,” he says.
“The logical progression from there is to integrate that data with a smart machine and even if it does nothing else but balance the quality of diagnostics, then that itself is a win. There have been some studies done in the US already where they've demonstrated that computer vision applications have a better rate of looking at tumour slides and saying ‘yes’ or ‘no’ than existing histology, oncology professors.”
The main barrier to AI advancing the healthcare space, however, is not technology, but money.
“We've solved a lot harder problems more quickly, it's just – particularly in the UK – the funding is not there or easily available,” McMullan says. “We'll see how that actually plays out in terms of engaging the NHS in that space. It's one of genuinely the easiest wins that we can actually make.”
The changing face of business AI
For businesses, advances in AI are driving considerable interest, but the time to deployment can prove an issue. However, companies are increasingly providing solutions to this problem.
Pure Storage has its own solution to this problem, in the form of AIRI and AIRI Mini. Developed in partnership with Nvidia, these products combine the previously disparate hardware and software components, dramatically reducing the deployment time in the process.
“We had a number of the joint customers, where they had some of their things, some of ours,” McMullan explains of the partnership. “We wanted to solve that whole ‘how do I make my data scientists effective on day one, not day 180?’ It wasn't a big risk to be honest; it was simply the obvious confluence of hardware leaders in two spaces which were generally considered quite hard.”
For businesses looking to deploy AI, this removed the need to focus on aspects of the technology beyond their own goals.
“The thing that we did I think that was best, which we don't talk enough about, is the software on top, the scaling toolkit, as our marketing team have called it,” he says.
“It links all of the Nvidia GPUs together so it just looks like one GPU to a data scientist. So he doesn't have to care about plumbing: the wiring, how many GPUs. You just throw that data model at it and, without any hyperbole, you can be up and in running in about two hours.”
“Unless you're an infrastructure person you don't want to be in the business of infrastructure.”
This rapid deployment, he says, is something companies have been asking for repeatedly, and which he sees great promise in due to its democratisation of the technology.
“If you can use a technology as simply as you can one of these,” he says, pointing to a smartphone, “then that interface, that ease of use, it is just two clicks and you're done. Unless you're an infrastructure person you don't want to be in the business of infrastructure.
“Data scientists don't know or care what the hardware is as long as it does what they want quickly, reliably. So the simplicity paradigm of one that we've been very strong on. Our founder is frankly ruthless on that whole thing. His thesis is if an eight-year-old can't do it then we should do it better.”
Continuous learning is essential
Of course while deployment time is seeing a dramatic drop, this does not mean that AI can become a static, unchanging resource.
“The machine always has to learn, it's like all of us. I think if you don't learn something every day you've kind of wasted your day, that's very much true with the intelligent machine: it always has to adapt, evolve, learn more, learn better,” he says.
“There's never a 'we've done AI, we've done ML'; it's a journey that will never end.”
The reason for this need for continuous training is that without it a model’s accuracy will decline.
“If you don't choose to engage with that community or the collaboration side of things, you can get to that point where your CIO can report to your board saying 'yes, we've done AI, we have our AI' but the model accuracy may be at 40% within three years if you don't pay attention,” he explains.
“If your model accuracy drops below 50% you may as well toss a coin, because you're going to get a better result. It's always about that continual iteration.”
Machines training machines: the missing piece of the AI puzzle
With so many advances in artificial intelligence, is there any aspect of the technology left to take it into the next phase? One, according to McMullan: the ability for machines to train machines.
“One of the challenges we have today is how to program the machine, or more accurately, how to explain what we do,” he explains.
“So facial recognition, we do that in a blink. What you have here is amazing in terms of image recognition, but we have no idea how that works. We emulate it with neural networks and hope it's kind of doing the same thing, but really we're not sure.
“I think for me the big step we're still missing is to have the machine train another machine. When you have that last capability of the machine making a smarter machine by software then that's the final thing to kick-start the hockey stick.”