Intelligence Quotient
A look at the latest in AI and how firms are applying it.
SAN FRANCISCO — It is time to get smart about artificial intelligence, which is evolving from Silicon Valley buzzword into a tool for brands.
And to do just that, hundreds of data scientists, engineers and academics huddled in the Hyatt Regency here for Re-work’s Deep Learning Summit, where experts from Apple, Google, Facebook, Amazon and elsewhere weighed in with the latest.
By its very definition, AI goes beyond mere smart technology to aim for intelligence, the sort that blurs the lines between what machines can do and what people still must. That dividing line is shifting by the day.
As attendees delved into the technical nuts and bolts — such as the Python programming language’s place in developing back- end AI infrastructure, how Generative Adversarial Networks can drive more humanlike thinking or ways cognitive behavior informs various models — an overarching sentiment prevailed: AI is not a single technology, and there’s no one preferred way to implement it.
The term has become a generic label covering fields such as machine-learning, natural language processing, computer vision, neural networks and others, and they often work together either linearly or in concert to produce desired results.
In retail, those results can take many forms. Think of the service bot that reduces the load on human workers, the voice assistant capable of responding to half-formed human requests based on context, and systems that can identify specific looks, who made them, where they’re available and for how much.
Product recommendations, long a staple in fashion and beauty e-commerce, become much more powerful when they match customers’ tastes.
To get there, companies like Luminoso employ artificial intelligence to understand consumer preferences for a range of clients, including retailers. “It looks like a word cloud, right?” said Bao Buli, an application engineer at Luminoso. He pulled up an example of work the company did for Sephora, starting with a jumble of words on a screen. Then he moved onto another page to reveal where the terms originated. “That actually comes from customer reviews. We can see the terms people use to describe things, and created connections between them. Why? So we can understand customer preferences.”
With that, Luminoso’s system builds associations between words like “dewy” and “glowy,” and it can see how often they’re used together and what other descriptors tend to follow. The result is more than a database of terms, but a context-aware system. When customers describe what they’re looking for, the company can get past the literal words people use to understand what they’re really looking for and make recommendations.
When it comes to collecting data, context is important, but when it comes to interpreting data and basing actions on it, context is critical.
Carl Vondrick, a research scientist at Google, explained how context traverses different media, from sound to visuals. Where they meet tends to be in video, and in one example, he pointed to a dog barking in a clip. The on-screen data flipped wildly, as the AI calculated the percentage chance of the dog being one breed over another based on his bark, and the computer vision saw more than a green backdrop. It identified grass and trees and figured the pooch was in or near a lawn or park.
AI algorithms are only as good as the quality and volume of data being fed into them, which works to Google’s advantage. “We can scale up visually using huge amounts of video,” Vondrick said. Indeed. The tech giant is owned by Alphabet, parent company of YouTube.
Of course, sound varies — from ambient noise to data gleaned from voice assistants as they style users — and the visuals can range from video shares of a dog playing outside to beauty vloggers to social media imagery shot at fashion week.
Conceptually, these efforts might sound like lab experiments, and for some organizations, that’s what they amount to. But for business, the work ultimately has to go beyond the lab. “You’ll see enterprise going out there snapping up as many Ph.D.s as possible,” said Maithili Mavinkurve, founder and chief executive officer of Sightline Innovation. “But does it help the enterprise? You have to look at the support and infrastructure — the software systems and networking — and you want to be able to quantify those objectives, so you can do it at scale.”
Companies and their development also should be clear about their goals, she said, and stay focused about the features they want to build.
The Canadian company specializes in machine-learning solutions for clients ranging from industrial to medical to commercial. For instance, for secondhand and vintage marketplace Skavnger, Sightline’s technology made it possible for buyers to find listings, without sellers worrying about crafting the perfect product description.
“It’s not just about making it [vaguely] better,” said Mavinkurve. “I call that the ‘ welcome to reality’ situation.”
Plenty of beauty and fashion brands seem to have gotten the memo. From innovators like Gap and L’Oréal to relative start-ups like ModiFace, Stitch Fix, Poshmark and many others, companies of all stripes are pursuing AI to drive better experiences for consumers.
Part of the reason for this push might be that every major tech company has doubled down on AI over the past year or two — including Salesforce, which spotlighted the technology at its Dreamforce conference. And they’re just getting started.
The other reason could be competitive pressure from Amazon, which has broad retail ambitions. Amazon’s AI efforts extend from its own Alexa voice platform to developer tools and much more. Last year, the company’s work hit a new threshold: Its technology became capable of fashion design, thanks to machine evaluations of images and algorithm-based decision-making.
The jury is still out on the actual merits of such work, but as machines get better at jobs previously thought to require the human touch, their value and uses go up. But so do the pitfalls.
The criminal system has started increasingly relying on artificial intelligence to determine sentencing and parole decisions. On the plus side, the approach can help bring more efficiencies to the penal system. But no technology is 100 percent foolproof, and when vendors protect their intellectual property by refusing to reveal the algorithms that determine the criminal risk of a given convict, the reality can start to border on dystopia.
Last year, Hong Kong-based Hanson Robotics’ Sophia became the first AI robot to gain citizenship, in Saudi Arabia.
If the goal of AI is to resemble human intelligence as much as possible, then the more the industry succeeds, the more disconcerting the result can be for the humans this tech was meant to serve.
These matters stand far apart from the retail bots and fashion recommendations designed to delight customers or help them. But they contribute to consumer perception and can create unease if not handled well. Is the communication and the methodology transparent? Is the data anonymized? How is it secured? Is participation opt-in or opt- out, and is that clear to the customer? Are the features promising enough value to make the effort worthwhile, and does it deliver on that promise?
Attention to these types of details shows dedication and care for the human consumers at the heart of AI, and must be as fundamental to the cause as its algorithms, data sets and categorizations.