Picture a dog. The image you have in your mind probably has a long history. It likely started when you were very young and came across a dog for the first time: Maybe you pet the dog, maybe it licked your hand. You met a few other dogs, and you got a sense of how they move, what they feel like, the space they occupy, what they sound like.

All these interactions make changes to your brain—specifically in the neocortex, the vast portion (about 70 percent) near the front of the skull that is responsible for giving us logic, reason, memory, and all those other higher functions that separate us from every other life form on the planet. (The “old” brain—that lower rear 30 percent—is responsible for more rote behaviors and drives: breathing, heartbeat, reflexes; anger, hunger, sex.) And this is how we learn—by moving through the world in three dimensions, actively connecting with our surroundings, and learning as we go.

That’s biological intelligence. But artificial intelligence—at least the way that it’s currently constructed—doesn’t work this way.

Let’s use Google as an example.

Google can identify a dog because the search engine has been fed millions of images of dogs—all labeled as such by humans—and it’s used those images to create a visual identification model based on things like shapes, lines, and colors. So if you ask Google what a dog is, Google Images will offer you billions of results. Google’s not the only one. From Microsoft and Amazon to countless buzzy Valley startups, when someone says they are involved in AI, they are generally using a massive stack of data to find patterns and build an algorithm based on that data.

But even with the seeming data dominance, the toddler who’s seen a few dogs has an advantage over the algorithm that has seen millions, says Donna Dubinsky (MBA 1981), CEO of Numenta, a Silicon Valley–based company she founded 13 years ago with her former Palm and Handspring cofounder, scientist Jeff Hawkins. “They would understand it in a much deeper way than that ten-millionth example of the Google Images search does. They know that a dog might be dangerous, that a dog might bark—they understand all these other characteristics about a dog, and they even generalize to an amazing extent. If they see a cartoon image of a purple dog, even though they’ve never seen a real purple dog in their life, they still know that it’s a dog. Whereas a Google engine may fail on that because it’s never seen a real purple dog—because there aren’t real purple dogs.”

That is what Numenta is after, says Dubinsky: recreating human intelligence by first understanding the complex way that humans learn—literally, the biological principles that guide our brains—and then translating those processes into code. Current AI is based on 50-year-old algorithms that have been supercharged with more data and computing power, she says. “It’s not that they aren’t accomplishing interesting things—they are,” she notes. “But they’re not doing it in a way that your brain does it. So the next big thing is going to be, How can we make computing more brain-like?”

And that, she says, is the key not only to better AI, but to a better understanding of what makes us human.

When Dubinsky declares something to be the next big thing, it isn’t empty Valley bravado.

She has worked at the forefront of personal computing at Apple, managing distribution and famously standing up to Steve Jobs’s plan to eliminate Apple’s US warehouses and decentralize distribution. (The details were captured in the 1986 HBS case, “Donna Dubinsky and Apple Computer, Inc.”) She and Hawkins then teamed up to usher in the era of handheld computing at Palm, where she was the founding CEO. When the duo partnered again at Handspring, they launched another foundational tech revolution with the smartphone.

But long before she redefined consumer technology, Dubinsky customized bowling shirts. When overcrowding at her Benton Harbor, Michigan, high school left her with a mornings-only class schedule, she spent the afternoons working retail at an athletic clothing shop. “I think it planted an entrepreneurial seed in me. I loved the hands-on aspect to it, I loved how you could see the results of what you do,” she says. After an “eye-opening” experience at Yale (“from small symbolic things to bigger, provocative, intellectual ideas—I just had none of that previously”), she worked as a banker in Philadelphia, funding mom-and-pop cable TV startups in rural Pennsylvania. It cemented her decision to apply to business school rather than law school, which was her original plan. “I came to the conclusion pretty quickly that I’d rather be on the business side than on the lawyer side because the lawyer was essentially a service arm to the business side, and the business side was really making the decisions,” says Dubinsky.

During a class at HBS, she watched as Dan Bricklin (MBA 1979) and Bob Frankston presented their VisiCalc invention—the first-ever electronic spreadsheet—and was blown away. Having spent countless hours retabulating forecasts with a calculator, one equation at a time, she immediately saw the value. “It was a lightbulb moment for me, really a classic moment of saying, ‘Every banker’s going to have one of those things,’ ” she says. VisiCalc was presented on an Apple computer, and Dubinsky—not knowing the difference between hardware and software back then—saw the machine’s logo and decided that’s where she wanted to work.

And while Apple did recruit on campus, it was by invite only, and Dubinsky didn’t make the list. Undeterred, she pleaded her case during the short window when the Apple recruiter left the interview room to get the next applicant, even bringing the recruiter a cup of coffee at one point. “I was really hamming it up,” she says. The recruiter eventually relented, and so began Dubinsky’s career of transforming the tech world.

The idea of Numenta—or at least the interest in its mission to understand the brain—was always hanging in the background of Hawkins and Dubinsky’s partnerships. In their very first meeting, when Hawkins was interviewing Dubinsky to be the CEO of Palm, he was frank. “Hey, look, I’m in this for four years. I really want to do something else—my passion’s in neuroscience and brain theory,” he told her. “The objective was—as soon as he could—to get back to working on the brain, and I had to buy into that to be his CEO,” says Dubinsky.

Numenta is the fulfillment of that promise. The company’s core technology is code based on Hawkins’s theories about brain function. (“We don’t do discovery by wet lab,” Hawkins clarifies. “We’re theorists.”) One of the first theories to become a business foundation was Hawkins’s finding that the brain’s neurons use their synapses to make predictions—about the future of patterns, the outcome of scenarios, expected behaviors—and that this is a critical element of human intelligence. Translated into code, these biological functions became a unique learning algorithm, which Numenta’s business partner Grok makes available as an anomaly detection tool, searching out odd signatures or disruptive patterns that could signify an attack or a system failure.

A fundamental component of the theory is how the brain represents information differently than computers do. Data entered into a computer—whether numbers or words—need to be translated into a machine language that the computer understands. The common language is ASCII, a binary code consisting of combinations of ones and zeros. But your brain doesn’t operate this way. Instead of ASCII, your brain understands sparse distributed representations (SDRs). SDRs are also zeros and ones, but they look very different from ASCII. An SDR is a very, very long string of mainly zeros and very few ones—and the ones have meaning. “If you look at the words ‘fir’ and ‘fire’ in ASCII text, you’d say they have a lot in common because their letters are almost entirely in common,” says Dubinsky. “But of course they have nothing in common in meaning. Now if you had the SDR for ‘fir’ and the SDR for ‘fire,’ they would have nothing in common—no ones that overlap—while the SDRs for ‘fire’ and ‘arson’ would have meaningful overlap.”

It’s further evidence of what Numenta is doing differently: They aren’t even speaking the same language as the other AI companies. And indeed, this SDR theory-turned-to-code has shown the ability to aid in natural language processing, with one of Numenta’s partners, Cortical.io, using it to help machines read lengthy, complex contracts to make sure they are in compliance.

But the business model has vacillated, a seesaw between a focus on research and a focus on business applications—everything from platform development to enterprise sales. David Yoffie, the Max and Doris Starr Professor of International Business Administration at HBS and a longtime friend of Dubinsky’s, highlights this tension in his 2016 case, “Numenta: Inventing and (or) Commercializing AI.” “The case itself describes five different paths that the company has taken. And every step along the way, they’ve had to make adjustments and try to figure out what was going to work and not work,” Yoffie says. “Ultimately, they’ve really been a 10-year experiment in scientific discovery of how to make AI into a potentially exciting new technology.”

“I think as a small company, I’m coming to the conclusion that you can’t do both research and business development at once,” says Dubinsky. In recent years, Numenta’s business model has focused on licensing and patents. The company’s research is open source, allowing academics and others to build on its work, but Numenta also offers standard commercial licenses—like the one Cortical.io is using—to pursue commercial opportunities. The idea behind making the work open source is both to accelerate the research and to offer transparency. “We don’t want it to be a big secret. So when people ask us, ‘What does it do?’ We don’t say, ‘It’s a big secret, we can’t tell you.’ We say, ‘Go look at it.’ ”

About a year and a half ago, though, Numenta began to shift its focus to research. Sample products that anyone could download to try out the tech were pulled off the website. “And we stopped trying to actively license our code,” says Numenta’s VP of engineering, Celeste Baranski. “This wasn’t a huge pivot but a tweak to concentrate completely on neuroscience research.” Her engineering group went from eight to two; she’s now working half-time.

The return to research was fueled by a breakthrough from the team—a new theory arguing that the cortex is not only processing all the information our senses are collecting but processing locations too. “Each little part of the cortex is figuring out where in the world things are and where they are relative to your body and where they are relative to each other,” says Hawkins. In this model, the brain is a powerful mapmaker, understanding, for instance, that these little black letters you are reading right now are not directly in front of your retina but a few feet away. It gives our world shape and dimension and structure in real time. “It was this other half of how the brain works that everyone had missed,” he says.

And so, Numenta became a different company. Yes, it meant a change in direction and all those disruptions and departures that can come with that, says Baranski. “But it’s the nature of what Numenta is doing.”

Most people assume that the human brain was solved long ago. That some scientist figured out how humans learn, published the findings in a white paper, and put it on a shelf, and it was settled. But that’s not the case. And that fact was shocking even 39 years ago when Jeff Hawkins came across a Scientific American article by Francis Crick—famed half of the DNA discovery duo Watson and Crick. “He wrote this powerful essay about how we have no idea how the brain works, and how we need a framework for thinking about it—and how everyone’s working on this problem, but no one has any idea what’s going on,” says Hawkins.

With that framework, he says, “it’s like a puzzle, and you don’t know how the puzzle pieces go together.” But thanks to recent discoveries, including that powerful discovery of the brain’s mapping activity, he believes that challenge Crick laid out almost four decades ago—and that Hawkins dutifully accepted—has been met. “We’ve discovered that framework,” he says.

Indeed, much of the energy on this late August day at Numenta is centered on this new framework, which will be presented in a paper and keynote speech the team is preparing for the annual summit of the Human Brain Project, a European neuroscience group. “It brings together a lot of bits and pieces that we’ve been working on into a coherent scaffold,” says Hawkins. “It helps orient us and helps us think about our research and where to go next.”

VP of Research Subutai Ahmad also employs the puzzle analogy to illuminate the road ahead. “Let’s say you have a 10,000-piece jigsaw puzzle. The first part is doing the outline, and filling in the first few pieces takes a long time. But the more you fill in, the faster you can fill in the rest of it, because there’s only so many places stuff can go. It feels like that here too.”

So what benefit does humanity get when we finally and fully understand how we learn? When we can comprehend the very core of what makes us human? What will it look like?

Hawkins demurs. “It’s not my strength, because it’s sort of like asking [computing pioneer John] von Neumann in 1937 what the computer was going to be,” he says. “If you were asking Einstein in 1918 what good is relativity theory, he’s not going to say, ‘Oh, we’re going to have these GPS satellites.’ He doesn’t know that. I can tell you broadly that I’m a believer that we’re going to build intelligent machines, and they’re going to be fantastically intelligent. I believe they are probably going to be important to the long-term survival of our species.”

Dubinsky is more specific. She sees potential applications in renewable energy and sustainability, for instance. “I think somebody taking this technology and applying it broadly in energy solutions has huge potential to help us manage climate change.” There are possibilities in genomics, she says, where data is expanding beyond our capacity to grasp it. There’s potential in health care, helping to eliminate simple things like false positives in hospital vital sign alarms. That very morning, she was listening to a podcast with a woman from the National Air and Space Museum talking about how we really need to send people to Mars to collect rocks and better understand evolution. “And I’m thinking, ‘Why can’t we send a robot to Mars to collect the rocks and bring them back?’ I think in robotics and areas like that it’s going to have to be technology based on our ideas.” They’re not there yet, though. “We can’t sit here and say, ‘Hey, today, somebody come and take our technology and program a robot to go to Mars.’ But our bottom line is that we’re on the path there, and the other people are not.”

At a research meeting at Numenta headquarters, Hawkins leads the team of neuroscientists in a long discussion about orientation—that is, how neurons determine where you are, and then predict where you’re going to be. Hawkins had pre-populated a whiteboard with a series of sketches that might as well have been Sanskrit to the uninitiated, and the conversation it sparks is equally impenetrable.

“Could the claustrum be represented by the ‘where’ region?”

“Does orientation get anchored like grid cells?”

“Is there any physiological data that tells us about these cells in level five?”

Hawkins conducts the conversation leaning against a window, cradling a coffee cup with two hands. He closes his eyes tightly when listening, and sometimes when he talks too—almost as if his own brain needs to quiet at least one input for complete focus.

The scene has all the trappings of Silicon Valley: hoodies, jeans, Mac laptops adorned with company logo stickers, themed conference room names. (We’re in Neuron, which sits next to Dendrite.) But the difference between this and the rest of the Valley is that no one here is building a product.

“The people here are here because they are very mission-driven,” says Dubinsky. “They want to work on this, they want to solve this problem.”

“A lot of people join a startup just because they want it to be the next Uber or Facebook or whatever. That’s not here,” says Baranski, whose career includes stints at HP, Palm, and Handspring, as well as founding her own startup. “People are not here to get rich tomorrow.”

Hawkins and Dubinsky lead Numenta with a balance they’ve honed over decades of working together—Hawkins the scientific visionary, Dubinsky the application wizard. As CEO, Dubinsky has a reputation for being approachable and transparent. She speaks directly and efficiently, with her Michigan accent mostly intact. Baranski shares a story about coming in to interview with Dubinsky at Handspring. “The receptionist goes, ‘Oh I think Donna’s in the stairwell, why don’t you go up there?’” Baranski says. “There was some mechanical problem in the stairwell or something, and she was there with her admin trying to fix whatever it was. It was so incredible the CEO would be there, working on the facilities.”

“When you talk to her, you feel like there are no secrets,” adds VP of Research Ahmad, who joined the company shortly after it was founded. “Really, whatever she might say about something is something she would say to everyone else—and she does say it to everyone else.”

Hawkins says he was struck by another of Dubinsky’s talents in their very first meeting, when he was interviewing her to be the CEO of Palm: her foresight. “She had a knack for understanding the importance of mobile computing when other people wouldn’t get it at that time,” he says. “It’s hard to imagine—you’ve got to go back to 1991—but the idea that everyone was going to have a computer in their pocket seemed incredibly stupid.”

It’s akin to the challenge Numenta and their biological vision for AI face today, says Dubinsky. Big shifts like this are jarring. She uses this analogy of the Rosetta Stone. “There was a whole field of scholarship about Egyptian hieroglyphics. And then they find the Rosetta Stone.” Overnight, the field—which had incorrectly accepted that the birds and pyramids represented birds and pyramids instead of the actual phonetic representations—was turned upside down. “All that previous scholarship was wrong,” she says. “You had to basically throw it out and turn in a different direction—and re-create dictionaries and re-create an understanding of these tombs and what they were and who was there—based on this whole new way of viewing things.”

“I think we, a little bit, do the same thing,” she says. She points to a whiteboard, where she had previously sketched out her view of the AI market: a cluster of black dots representing the mass of other AI companies using biological-sounding terms—deep learning, neural networks—and a few inches away, a single black dot representing Numenta. “We’re going over here, an entirely different direction, and those people who are going over there can’t go with us.”

“It’s the next people,” she says. “They can go with us.”

But getting there isn’t just a matter of getting the science right. If Dubinsky and Numenta want to start a revolution, it’s going to take some groundwork. And that’s top of mind: During a morning standup meeting, team members run through a series of updates on journal articles, blog posts, and podcasts. “The work is extremely exciting—we think it’s breakthrough—but it is very hard to figure out how to package it in a way that is digestible by normal people,” she says. “One of the examples I like to use is the whole notion of natural selection and evolution. Well, wow, how did a big idea like that take hold? A lot of people were very skeptical at first: ‘What do you mean? Evolution is just random? That can’t be.’ ”

But revolution is a comfortable challenge for Dubinsky. “I’m not outside my depth here,” she says. This is what she’s done her whole career: figure out how to explain revolutionary technology to a disbelieving crowd—and convince the world that they’ve been reading the hieroglyphics all wrong.

Post a Comment