Maybe Ed Doran really does like to listen to Jamaican-infused Nordic metal bands, maybe he doesn’t. Either way, there are times he wants to crank the music up to 11, and times it’s better muted.
Unfortunately, artificial intelligence can’t always tell the difference, and that missing context ranks among the key challenges developers face as they try to craft a smoother and more-useful interaction between humans and machines, the principal program manager at Microsoft Research AI said at South by Southwest earlier this month.
Despite the occasional bold prediction or prophesy of doom, the discussions about AI at this year’s SXSW festivities took a more pragmatic tone this year than in the past.
While several panels still contemplated our long-term future — one, for example, considered the possibility of consciousness in machines — most of the researchers, developers and pundits set their sights on the challenges AI faces in the here and now.
They talked about improving human-machine interactions, reducing bias in data sets and developer teams and, as Doran noted, enhancing an AI system’s sense of context and human environment.
“We’re trying to get a much more nuanced view of the personas we carry throughout the day,” Doran said of his Microsoft AI research team. “I might like listening to Icelandic death reggae in my office, but not when my boss comes in.”
Consider, for example, one of Doran’s developers, who often could be found at a neighborhood bar around midnight. Because he was there often enough at that time of night, his AI-powered assistant assumed that the bar was his home.
The infraction seems harmless enough, but it underscores why Doran and his teams adopted an approach to AI development that’s designed to put more control in the user’s hands. In essence, Microsoft’s Cortana AI assistant might suggest all it knows about the user, and then let the user note what’s right or wrong in that information.
This could help the system better “know” human users and their preferences. And even if users mislead it, intentionally or not, they remain in control. One developer even suggested a “Vegas mode,” Doran said, allowing users to go back and erase entire blocks of time so the AI couldn’t process and learn from the data gathered during that stretch.
But the ability of a machine to more deeply understand the relationships between people — and between people and the world they inhabit — will require AI systems that can experience the world more like humans do, said Nell Watson, an AI expert and futurist on the faculty of Singularity University.
Thinking machines will take “a leap forward” once developers can embed machine intelligence within that sort of environment, she said.
Adam Cheyer, the co-founder of Viv Labs and one of Siri’s creators, joined Watson on an “Exploring Innovations in AI” panel and shared her sense that significant breakthroughs were coming in the near future. Cheyer said most people will use an AI-powered assistant for almost everything they currently use the Web for, and it will happen within the next two to three years.
And as more people employ AI assistants and developers install greater processing and technical capability into them, he said, they will coordinate increasingly complex tasks and help users navigate the wealth of information around them.
A world of ‘super stimuli’
Watson cautioned, however, that we might not always like what we learn.
Already, one AI model known as generative adversarial networks, or GANs, enables the creation of deepfakes – machine-produced images and videos that, to the human eye, appear original and authentic.
The model works by pitting two systems against each other, with the first trying to develop something to fool the second, and the second trying to catch the first at its game – and both honing their capabilities as they go.
One need not think too hard to imagine ways nefarious actors could use deepfakes to mislead people. But even if not used to mislead, these sorts of systems could lead toward a world of “super stimuli,” Watson said, producing the unending AI versions of “the juiciest hamburger.” We’d have so many super stimuli, reality would pale in comparison.
“My fear is that machines will be like a mirror on the wall that tells us we’re not the most beautiful creature in the world,” she said.
Bias already exists in systems with far more impact, including predictive policing systems and technologies that recommend bail and probation terms for criminals. Throughout many of the AI panels during SXSW, developers and researchers stressed the need for greater consideration of the obvious and unseen biases in existing data sets.
To some extent, the concerns might not be so dire if AI systems could explain themselves. Currently, thinking machines can do tremendous, even super-human, feats on narrowly targeted functions, but they can’t yet explain why they come to the decisions they make. Oftentimes, the process inside the black box is so complex, human minds couldn’t track it even if they could peek inside.
So until AI agents can better explain their results, developers and companies will need to remain aware of the obvious and unnoticed biases that exist in their data sets.
“The world will look far different 10 years from now,” said Jeff Nelson, the founder and CEO of Cinchapi. “So if all these formative systems are created by monochromatic people … you’ll have a bunch of technologies that most of the world doesn’t embrace because those technologies don’t embrace them.”
‘Such a young field’
Fei Fei Li and Megan Smith are among the experts leading the charge on diversity and countering bias in teams and data. Li is a Stanford University computer science professor currently working with Google, and Smith is a former CTO at the White House, now heading up Shift7, a startup looking to promote more inclusive participation in high-tech.
In their panel, Smith noted the huge imbalance in dialog spoken by male and female actors. In a 2016 study of 2,000 movie scripts, 75 percent had male actors speaking at least 60 percent of the dialog and often far more.
Whatever the excuse for the discrepancy, it represents a misleading sample of dialog as it’s used in the real world. So, employing movies as data sets to train a language-processing system isn’t entirely unlike a foreigner watching John Wayne movies to learn how to speak English.
Li cuts the AI world a little bit of slack.
“It’s such a young field, really only about 16 years old, but one of the most profound pursuits humanity has created,” she said.
But she’ll only cut it a little bit of slack. To bring beneficial AI technology to the people, she said, the development community needs “a sustained pipeline” of new technologists and thinkers across a range disciplines.
And, Smith said, they need to get their hands dirty. After all, she said, “You don’t go to (physical education) and the teacher says, ‘Sit down and open your books. We’re going to do baseball today.’”
News on Open Source is free and unlimited. Access to the rest of 512tech.com comes with an American-Statesman digital subscription, which also includes myStatesman.com and the ePaper edition. Subscribe at statesman.com/subscribe.