Saffron Technology: The AI company quietly changing the world

It’s very easy to get lost in the potential of artificial intelligence when talking to Saffron Technology’s Gayle Sheppard. Fortunately, 31 minutes in, her iPhone reminded me exactly how far commercial AI is from primetime. Tucked away in her bag in the corner of the meeting room, Siri’s virtual ears prick up the first time her name is dropped into our conversation. “She thinks I insulted her,” Sheppard quips, after Siri gives up with an “I’m sorry, I didn’t get that,” before piping down and returning to silent eavesdropping.

While it’s not exactly an exclusive to reveal that Siri isn’t the best showcase of how brilliant AI can be, the juxtaposition is quite amusing. Here I am talking to Gayle Sheppard about how her company, acquired by Intel in 2015 after 16 years of independence, is quietly changing the world in subtle but undeniable ways, and a flashy upstart can’t resist trying to grab the spotlight.

What do I mean by “changing the world”? Sheppard highlights three examples of Saffron’s work that couldn’t be further apart: accurately predicting when aeroplane parts will fail; predicting future patterns in fraud and money laundering; and telling the difference between restrictive cardiomyopathy and constrictive pericarditis.

It’s the last of these that is the most instructive of how transformative the system can be. Saffron was contacted by a New York cardiologist – Dr Partho Sengupta – who believed that Saffron’s technology might be able to improve condition diagnosis. Even to an expert human eye, echocardiograms of the two conditions look very similar – Sengupta’s own eye was right around 76% of the time, but less experienced doctors were averaging just over 50%.

The doctors were basing their analyses on around seven different attributes, but there were many more that weren’t getting a look in. A mind-boggling amount, in fact. “Twenty times in a single heartbeat we capture something like 10,000 attributes in six different zones with 19 different metrics,” Sheppard explains. Using the philosophy that anything can matter, Saffron’s artificial intelligence was able to push the correct diagnosis up to 96% in just two months. “A lA lot of data, unrestrained: no attributes pre-selected. That’s where AI can really help the common good.”

That example really sums up Saffron’s approach to AI. The company is an example of “lazy learning AI” – meaning that it learns everything it can, and then works backwards from there. “The role of experience is rich and robust, and when you try to reduce it down to a few attributes you lose a lot of information and a lot of personalisation,” Sheppard explains. “The fundamentals are: learn about everything because anything can matter.”

Unlike deep learning systems, Saffron Technology has no black box, and the approach is completely transparent. “AI has to be explanatory or we lose confidence – we have no trust,” Sheppard explains. “The whole idea is that we don’t know what we don’t know, and therefore we can only imagine and model things with either our imagination or what we know.

“So when we’re constrained to models, we’re always going to leave something out. We’re always going to reduce the accuracy, inevitably.” That’s not to say that deep learning is necessarily inferior to the experiential learning Saffron offers. “When we’re in deep learning, models are very important. We need to know what you look like. I need to make sure that when I identify you, I’m not picking someone who looks like you.”

Sheppard speaks with the authority of someone who has worked in the industry for nearly two decades. “In 1999 – good idea, but unstructured data wasn’t well understood yet,” she explains. There were a few bumps in the road following the dotcom bubble burst and the September 11 attacks, but Saffron has lasted the pace and now finds itself part of Intel. “They understood what was possible with Saffron, and they’ve been preparing an AI portfolio for some time,” explains Sheppard, who regards the two companies as a “really good fit”, and is keen to tap into the company’s well-regarded community of developers.

Any developers dipping their toes into AI will find things much more welcoming than they were in 1999. Customers now have much more data to share, meaning that Saffron can be a lot more helpful. “The customers are really smart – they have data scientists and machine-learning experts. We go in and we’re not educating as much – they already have a great foundation of knowledge to work with, and it’s really becoming a marriage of the minds.”

Of course many of those companies have invested in data scientists and machine-learning experts for the same reasons that people fear AI: because it can reduce headcount. I ask Sheppard where Saffron stands on people’s fear of redundancy. “We’re not working on any problems at the present time, proposed to us by industry, that are job-eliminating,” Sheppard says. “We’re focusing on jobs that can’t get done very easily by humans.” Take the area of fraud investigation, for example: “You can’t hire enough people to investigate everything – it’s just impossible because of the volume of transactions. It’s really about getting everything done, so there’s less stress on the poor investigators – I can’t imagine being in that role, because there’s some serious fines imposed on the banks if they don’t get these things investigated in 90 days!”

When I think about jobs that can’t practically be done by humans, I don’t think of fraud analysis or remote well platforms in the North Sea: I think of online harassment. Facebook, Twitter and Google have no solution to trolling and abuse, because it’s just too much for a human team to handle. Could AI step in? “That’s a great idea – we haven’t applied our technology to that yet. Intel announced an initiative in anti-harassment which we haven’t yet been involved in – you’ve just motivated me to ping that person an email to say ‘come on, let’s go. Let’s see how we can help you’.” So if harassment on the internet is fixed by the time you read this, you’ll know who to thank.

It’s pretty clear from talking to Sheppard that the problems that appeal the most are the ones humans just don’t have the capability to work on (“I sort of feel boring at the moment, because I’m just solving some real-world mundane problems,” she says at one point), but those aren’t always ones that fit with our profit-driven corporate world.

“This is really important,” she explains, emphasising that whenever large corporations are involved in AI – including Intel – that they’re doing so “for mankind and bettering the world, as opposed to just for economic gain”. Whether or not companies pay attention to those words over the coming decades could decide what our world looks like in 2050 – whether or not humans end up sidelined.