Archiv der Kategorie: Machine Learning

The Pentagon’s Push to Program Soldiers’ Brains

The military wants future super-soldiers to control robots with their thoughts.

I. Who Could Object?

“Tonight I would like to share with you an idea that I am extremely passionate about,” the young man said. His long black hair was swept back like a rock star’s, or a gangster’s. “Think about this,” he continued. “Throughout all human history, the way that we have expressed our intent, the way we have expressed our goals, the way we have expressed our desires, has been limited by our bodies.” When he inhaled, his rib cage expanded and filled out the fabric of his shirt. Gesturing toward his body, he said, “We are born into this world with this. Whatever nature or luck has given us.“

His speech then took a turn: “Now, we’ve had a lot of interesting tools over the years, but fundamentally the way that we work with those tools is through our bodies.” Then a further turn: “Here’s a situation that I know all of you know very well—your frustration with your smartphones, right? This is another tool, right? And we are still communicating with these tools through our bodies.”

And then it made a leap: “I would claim to you that these tools are not so smart. And maybe one of the reasons why they’re not so smart is because they’re not connected to our brains. Maybe if we could hook those devices into our brains, they could have some idea of what our goals are, what our intent is, and what our frustration is.”

So began “Beyond Bionics,” a talk by Justin C. Sanchez, then an associate professor of biomedical engineering and neuroscience at the University of Miami, and a faculty member of the Miami Project to Cure Paralysis. He was speaking at a tedx conference in Florida in 2012. What lies beyond bionics? Sanchez described his work as trying to “understand the neural code,” which would involve putting “very fine microwire electrodes”—the diameter of a human hair—“into the brain.” When we do that, he said, we would be able to “listen in to the music of the brain” and “listen in to what somebody’s motor intent might be” and get a glimpse of “your goals and your rewards” and then “start to understand how the brain encodes behavior.”

He explained, “With all of this knowledge, what we’re trying to do is build new medical devices, new implantable chips for the body that can be encoded or programmed with all of these different aspects. Now, you may be wondering, what are we going to do with those chips? Well, the first recipients of these kinds of technologies will be the paralyzed. It would make me so happy by the end of my career if I could help get somebody out of their wheelchair.”

Sanchez went on, “The people that we are trying to help should never be imprisoned by their bodies. And today we can design technologies that can help liberate them from that. I’m truly inspired by that. It drives me every day when I wake up and get out of bed. Thank you so much.” He blew a kiss to the audience.

A year later, Justin Sanchez went to work for the Defense Advanced Research Projects Agency, the Pentagon’s R&D department. At darpa, he now oversees all research on the healing and enhancement of the human mind and body. And his ambition involves more than helping get disabled people out of their wheelchair—much more.

DARPA has dreamed for decades of merging human beings and machines. Some years ago, when the prospect of mind-controlled weapons became a public-relations liability for the agency, officials resorted to characteristic ingenuity. They recast the stated purpose of their neurotechnology research to focus ostensibly on the narrow goal of healing injury and curing illness. The work wasn’t about weaponry or warfare, agency officials claimed. It was about therapy and health care. Who could object? But even if this claim were true, such changes would have extensive ethical, social, and metaphysical implications. Within decades, neurotechnology could cause social disruption on a scale that would make smartphones and the internet look like gentle ripples on the pond of history.

Most unsettling, neurotechnology confounds age-old answers to this question: What is a human being?

II. High Risk, High Reward

In his 1958 State of the Union address, President Dwight Eisenhower declared that the United States of America “must be forward-looking in our research and development to anticipate the unimagined weapons of the future.” A few weeks later, his administration created the Advanced Research Projects Agency, a bureaucratically independent body that reported to the secretary of defense. This move had been prompted by the Soviet launch of the Sputnik satellite. The agency’s original remit was to hasten America’s entry into space.

During the next few years, arpa’s mission grew to encompass research into “man-computer symbiosis” and a classified program of experiments in mind control that was code-named Project Pandora. There were bizarre efforts that involved trying to move objects at a distance by means of thought alone. In 1972, with an increment of candor, the word Defense was added to the name, and the agency became darpa. Pursuing its mission, darpa funded researchers who helped invent technologies that changed the nature of battle (stealth aircraft, drones) and shaped daily life for billions (voice-recognition technology, GPS devices). Its best-known creation is the internet.

The agency’s penchant for what it calls “high-risk, high-reward” research ensured that it would also fund a cavalcade of folly. Project Seesaw, a quintessential Cold War boondoggle, envisioned a “particle-beam weapon” that could be deployed in the event of a Soviet attack. The idea was to set off a series of nuclear explosions beneath the Great Lakes, creating a giant underground chamber. Then the lakes would be drained, in a period of 15 minutes, to generate the electricity needed to set off a particle beam. The beam would accelerate through tunnels hundreds of miles long (also carved out by underground nuclear explosions) in order to muster enough force to shoot up into the atmosphere and knock incoming Soviet missiles out of the sky. During the Vietnam War, darpa tried to build a Cybernetic Anthropomorphous Machine, a jungle vehicle that officials called a “mechanical elephant.”

The diverse and sometimes even opposing goals of darpa scientists and their Defense Department overlords merged into a murky, symbiotic research culture—“unencumbered by the typical bureaucratic oversight and uninhibited by the restraints of scientific peer review,” Sharon Weinberger wrote in a recent book, The Imagineers of War. In Weinberger’s account, darpa’s institutional history involves many episodes of introducing a new technology in the context of one appealing application, while hiding other genuine but more troubling motives. At darpa, the left hand knows, and doesn’t know, what the right hand is doing.

The agency is deceptively compact. A mere 220 employees, supported by about 1,000 contractors, report for work each day at darpa’s headquarters, a nondescript glass-and-steel building in Arlington, Virginia, across the street from the practice rink for the Washington Capitals. About 100 of these employees are program managers—scientists and engineers, part of whose job is to oversee about 2,000 outsourcing arrangements with corporations, universities, and government labs. The effective workforce of darpa actually runs into the range of tens of thousands. The budget is officially said to be about $3 billion, and has stood at roughly that level for an implausibly long time—the past 14 years.

The Biological Technologies Office, created in 2014, is the newest of darpa’s six main divisions. This is the office headed by Justin Sanchez. One purpose of the office is to “restore and maintain warfighter abilities” by various means, including many that emphasize neurotechnology—applying engineering principles to the biology of the nervous system. For instance, the Restoring Active Memory program develops neuroprosthetics—tiny electronic components implanted in brain tissue—that aim to alter memory formation so as to counteract traumatic brain injury. Does darpa also run secret biological programs? In the past, the Department of Defense has done such things. It has conducted tests on human subjects that were questionable, unethical, or, many have argued, illegal. The Big Boy protocol, for example, compared radiation exposure of sailors who worked above and below deck on a battleship, never informing the sailors that they were part of an experiment.

Eddie Guy
Last year I asked Sanchez directly whether any of darpa’s neurotechnology work, specifically, was classified. He broke eye contact and said, “I can’t—We’ll have to get off that topic, because I can’t answer one way or another.” When I framed the question personally—“Are you involved with any classified neuroscience project?”—he looked me in the eye and said, “I’m not doing any classified work on the neurotechnology end.”

If his speech is careful, it is not spare. Sanchez has appeared at public events with some frequency (videos are posted on darpa’s YouTube channel), to articulate joyful streams of good news about darpa’s proven applications—for instance, brain-controlled prosthetic arms for soldiers who have lost limbs. Occasionally he also mentions some of his more distant aspirations. One of them is the ability, via computer, to transfer knowledge and thoughts from one person’s mind to another’s.

III. “We Try to Find Ways to Say Yes”

Medicine and biology were of minor interest to darpa until the 1990s, when biological weapons became a threat to U.S. national security. The agency made a significant investment in biology in 1997, when darpa created the Controlled Biological Systems program. The zoologist Alan S. Rudolph managed this sprawling effort to integrate the built world with the natural world. As he explained it to me, the aim was “to increase, if you will, the baud rate, or the cross-communication, between living and nonliving systems.” He spent his days working through questions such as “Could we unlock the signals in the brain associated with movement in order to allow you to control something outside your body, like a prosthetic leg or an arm, a robot, a smart home—or to send the signal to somebody else and have them receive it?”

Human enhancement became an agency priority. “Soldiers having no physical, physiological, or cognitive limitation will be key to survival and operational dominance in the future,” predicted Michael Goldblatt, who had been the science and technology officer at McDonald’s before joining darpa in 1999. To enlarge humanity’s capacity to “control evolution,” he assembled a portfolio of programs with names that sounded like they’d been taken from video games or sci-fi movies: Metabolic Dominance, Persistence in Combat, Continuous Assisted Performance, Augmented Cognition, Peak Soldier Performance, Brain-Machine Interface.

The programs of this era, as described by Annie Jacobsen in her 2015 book, The Pentagon’s Brain, often shaded into mad-scientist territory. The Continuous Assisted Performance project attempted to create a “24/7 soldier” who could go without sleep for up to a week. (“My measure of success,” one darpa official said of these programs, “is that the International Olympic Committee bans everything we do.”)

Dick Cheney relished this kind of research. In the summer of 2001, an array of “super-soldier” programs was presented to the vice president. His enthusiasm contributed to the latitude that President George W. Bush’s administration gave darpa—at a time when the agency’s foundation was shifting. Academic science gave way to tech-industry “innovation.” Tony Tether, who had spent his career working alternately for Big Tech, defense contractors, and the Pentagon, became darpa’s director. After the 9/11 attacks, the agency announced plans for a surveillance program called Total Information Awareness, whose logo included an all-seeing eye emitting rays of light that scanned the globe. The pushback was intense, and Congress took darpa to task for Orwellian overreach. The head of the program—Admiral John Poindexter, who had been tainted by scandal back in the Reagan years—later resigned, in 2003. The controversy also drew unwanted attention to darpa’s research on super-soldiers and the melding of mind and machine. That research made people nervous, and Alan Rudolph, too, found himself on the way out.

In this time of crisis, darpa invited Geoff Ling, a neurology‑ICU physician and, at the time, an active-duty Army officer, to join the Defense Sciences Office. (Ling went on to work in the Biological Technologies Office when it spun out from Defense Sciences, in 2014.) When Ling was interviewed for his first job at darpa, in 2002, he was preparing for deployment to Afghanistan and thinking about very specific combat needs. One was a “pharmacy on demand” that would eliminate the bulk of powdery fillers from drugs in pill or capsule form and instead would formulate active ingredients for ingestion via a lighter, more compact, dissolving substance—like Listerine breath strips. This eventually became a darpa program. The agency’s brazen sense of possibility buoyed Ling, who recalls with pleasure how colleagues told him, “We try to find ways to say yes, not ways to say no.” With Rudolph gone, Ling picked up the torch.

Ling talks fast. He has a tough-guy voice. The faster he talks, the tougher he sounds, and when I met him, his voice hit top speed as he described a first principle of Defense Sciences. He said he had learned this “particularly” from Alan Rudolph: “Your brain tells your hands what to do. Your hands basically are its tools, okay? And that was a revelation to me.” He continued, “We are tool users—that’s what humans are. A human wants to fly, he builds an airplane and flies. A human wants to have recorded history, and he creates a pen. Everything we do is because we use tools, right? And the ultimate tools are our hands and feet. Our hands allow us to work with the environment to do stuff, and our feet take us where our brain wants to go. The brain is the most important thing.”

Ling connected this idea of the brain’s primacy with his own clinical experience of the battlefield. He asked himself, “How can I liberate mankind from the limitations of the body?” The program for which Ling became best known is called Revolutionizing Prosthetics. Since the Civil War, as Ling has said, the prosthetic arm given to most amputees has been barely more sophisticated than “a hook,” and not without risks: “Try taking care of your morning ablutions with that bad boy, and you’re going to need a proctologist every goddamn day.” With help from darpa colleagues and academic and corporate researchers, Ling and his team built something that was once all but unimaginable: a brain-controlled prosthetic arm.

No invention since the internet has been such a reliable source of good publicity for darpa. Milestones in its development were hailed with wonder. In 2012, 60 Minutes showed a paralyzed woman named Jan Scheuermann feeding herself a bar of chocolate using a robotic arm that she manipulated by means of a brain implant.

​Eddie Guy
Yet darpa’s work to repair damaged bodies was merely a marker on a road to somewhere else. The agency has always had a larger mission, and in a 2015 presentation, one program manager—a Silicon Valley recruit—described that mission: to “free the mind from the limitations of even healthy bodies.” What the agency learns from healing makes way for enhancement. The mission is to make human beings something other than what we are, with powers beyond the ones we’re born with and beyond the ones we can organically attain.

The internal workings of darpa are complicated. The goals and values of its research shift and evolve in the manner of a strange, half-conscious shell game. The line between healing and enhancement blurs. And no one should lose sight of the fact that D is the first letter in darpa’s name. A year and a half after the video of Jan Scheuermann feeding herself chocolate was shown on television, darpa made another video of her, in which her brain-computer interface was connected to an F-35 flight simulator, and she was flying the airplane. darpa later disclosed this at a conference called Future of War.

Geoff Ling’s efforts have been carried on by Justin Sanchez. In 2016, Sanchez appeared at darpa’s “Demo Day” with a man named Johnny Matheny, whom agency officials describe as the first “osseointegrated” upper-limb amputee—the first man with a prosthetic arm attached directly to bone. Matheny demonstrated what was, at the time, darpa’s most advanced prosthetic arm. He told the attendees, “I can sit here and curl a 45-pound dumbbell all day long, till the battery runs dead.” The next day, Gizmodo ran this headline above its report from the event: “darpa’s Mind-Controlled Arm Will Make You Wish You Were a Cyborg.”

Since then, darpa’s work in neurotechnology has avowedly widened in scope, to embrace “the broader aspects of life,” Sanchez told me, “beyond the person in the hospital who is using it to heal.” The logical progression of all this research is the creation of human beings who are ever more perfect, by certain technological standards. New and improved soldiers are necessary and desirable for darpa, but they are just the window-display version of the life that lies ahead.

IV. “Over the Horizon”

Consider memory, Sanchez told me: “Everybody thinks about what it would be like to give memory a boost by 20, 30, 40 percent—pick your favorite number—and how that would be transformative.” He spoke of memory enhancement through neural interface as an alternative form of education. “School in its most fundamental form is a technology that we have developed as a society to help our brains to do more,” he said. “In a different way, neurotechnology uses other tools and techniques to help our brains be the best that they can be.” One technique was described in a 2013 paper, a study involving researchers at Wake Forest University, the University of Southern California, and the University of Kentucky. Researchers performed surgery on 11 rats. Into each rat’s brain, an electronic array—featuring 16 stainless-steel wires—was implanted. After the rats recovered from surgery, they were separated into two groups, and they spent a period of weeks getting educated, though one group was educated more than the other.

The less educated group learned a simple task, involving how to procure a droplet of water. The more educated group learned a complex version of that same task—to procure the water, these rats had to persistently poke levers with their nose despite confounding delays in the delivery of the water droplet. When the more educated group of rats attained mastery of this task, the researchers exported the neural-firing patterns recorded in the rats’ brains—the memory of how to perform the complex task—to a computer.

“What we did then was we took those signals and we gave it to an animal that was stupid,” Geoff Ling said at a darpa event in 2015—meaning that researchers took the neural-firing patterns encoding the memory of how to perform the more complex task, recorded from the brains of the more educated rats, and transferred those patterns into the brains of the less educated rats—“and that stupid animal got it. They were able to execute that full thing.” Ling summarized: “For this rat, we reduced the learning period from eight weeks down to seconds.”

“They could inject memory using the precise neural codes for certain skills,” Sanchez told me. He believes that the Wake Forest experiment amounts to a foundational step toward “memory prosthesis.” This is the stuff of The Matrix. Though many researchers question the findings—cautioning that, really, it can’t be this simple—Sanchez is confident: “If I know the neural codes in one individual, could I give that neural code to another person? I think you could.” Under Sanchez, darpa has funded human experiments at Wake Forest, the University of Southern California, and the University of Pennsylvania, using similar mechanisms in analogous parts of the brain. These experiments did not transfer memory from one person to another, but instead gave individuals a memory “boost.” Implanted electrodes recorded neuronal activity associated with recognizing patterns (at Wake Forest and USC) and memorizing word lists (at Penn) in certain brain circuits. Then electrodes fed back those recordings of neuronal activity into the same circuits as a form of reinforcement. The result, in both cases, was significantly improved memory recall.

Doug Weber, a neural engineer at the University of Pittsburgh who recently finished a four-year term as a darpa program manager, working with Sanchez, is a memory-transfer skeptic. Born in Wisconsin, he has the demeanor of a sitcom dad: not too polished, not too rumpled. “I don’t believe in the infinite limits of technology evolution,” he told me. “I do believe there are going to be some technical challenges which are impossible to achieve.” For instance, when scientists put electrodes in the brain, those devices eventually fail—after a few months or a few years. The most intractable problem is blood leakage. When foreign material is put into the brain, Weber said, “you undergo this process of wounding, bleeding, healing, wounding, bleeding, healing, and whenever blood leaks into the brain compartment, the activity in the cells goes way down, so they become sick, essentially.” More effectively than any fortress, the brain rejects invasion.

Even if the interface problems that limit us now didn’t exist, Weber went on to say, he still would not believe that neuroscientists could enable the memory-prosthesis scenario. Some people like to think about the brain as if it were a computer, Weber explained, “where information goes from A to B to C, like everything is very modular. And certainly there is clear modular organization in the brain. But it’s not nearly as sharp as it is in a computer. All information is everywhere all the time, right? It’s so widely distributed that achieving that level of integration with the brain is far out of reach right now.”

Peripheral nerves, by contrast, conduct signals in a more modular fashion. The biggest, longest peripheral nerve is the vagus. It connects the brain with the heart, the lungs, the digestive tract, and more. Neuroscientists understand the brain’s relationship with the vagus nerve more clearly than they understand the intricacies of memory formation and recall among neurons within the brain. Weber believes that it may be possible to stimulate the vagus nerve in ways that enhance the process of learning—not by transferring experiential memories, but by sharpening the facility for certain skills.

To test this hypothesis, Weber directed the creation of a new program in the Biological Technologies Office, called Targeted Neuroplasticity Training (TNT). Teams of researchers at seven universities are investigating whether vagal-nerve stimulation can enhance learning in three areas: marksmanship, surveillance and reconnaissance, and language. The team at Arizona State has an ethicist on staff whose job, according to Weber, “is to be looking over the horizon to anticipate potential challenges and conflicts that may arise” regarding the ethical dimensions of the program’s technology, “before we let the genie out of the bottle.” At a TNT kickoff meeting, the research teams spent 90 minutes discussing the ethical questions involved in their work—the start of a fraught conversation that will broaden to include many others, and last for a very long time.

DARPA officials refer to the potential consequences of neurotechnology by invoking the acronym elsi, a term of art devised for the Human Genome Project. It stands for “ethical, legal, social implications.” The man who led the discussion on ethics among the research teams was Steven Hyman, a neuroscientist and neuroethicist at MIT and Harvard’s Broad Institute. Hyman is also a former head of the National Institute of Mental Health. When I spoke with him about his work on darpa programs, he noted that one issue needing attention is “cross talk.” A man-machine interface that does not just “read” someone’s brain but also “writes into” someone’s brain would almost certainly create “cross talk between those circuits which we are targeting and the circuits which are engaged in what we might call social and moral emotions,” he said. It is impossible to predict the effects of such cross talk on “the conduct of war” (the example he gave), much less, of course, on ordinary life.

Weber and a darpa spokesperson related some of the questions the researchers asked in their ethics discussion: Who will decide how this technology gets used? Would a superior be able to force subordinates to use it? Will genetic tests be able to determine how responsive someone would be to targeted neuroplasticity training? Would such tests be voluntary or mandatory? Could the results of such tests lead to discrimination in school admissions or employment? What if the technology affects moral or emotional cognition—our ability to tell right from wrong or to control our own behavior?

Recalling the ethics discussion, Weber told me, “The main thing I remember is that we ran out of time.”

V. “You Can Weaponize Anything”

In The Pentagon’s Brain, Annie Jacobsen suggested that darpa’s neurotechnology research, including upper-limb prosthetics and the brain-machine interface, is not what it seems: “It is likely that darpa’s primary goal in advancing prosthetics is to give robots, not men, better arms and hands.” Geoff Ling rejected the gist of her conclusion when I summarized it for him (he hadn’t read the book). He told me, “When we talk about stuff like this, and people are looking for nefarious things, I always say to them, ‘Do you honestly believe that the military that your grandfather served in, your uncle served in, has changed into being Nazis or the Russian army?’ Everything we did in the Revolutionizing Prosthetics program—everything we did—is published. If we were really building an autonomous-weapons system, why would we publish it in the open literature for our adversaries to read? We hid nothing. We hid not a thing. And you know what? That meant that we didn’t just do it for America. We did it for the world.”

I started to say that publishing this research would not prevent its being misused. But the terms use and misuse overlook a bigger issue at the core of any meaningful neurotechnology-ethics discussion. Will an enhanced human being—a human being possessing a neural interface with a computer—still be human, as people have experienced humanity through all of time? Or will such a person be a different sort of creature?

​Eddie Guy
The U.S. government has put limits on darpa’s power to experiment with enhancing human capabilities. Ling says colleagues told him of a “directive”: “Congress was very specific,” he said. “They don’t want us to build a superperson.” This can’t be the announced goal, Congress seems to be saying, but if we get there by accident—well, that’s another story. Ling’s imagination remains at large. He told me, “If I gave you a third eye, and the eye can see in the ultraviolet, that would be incorporated into everything that you do. If I gave you a third ear that could hear at a very high frequency, like a bat or like a snake, then you would incorporate all those senses into your experience and you would use that to your advantage. If you can see at night, you’re better than the person who can’t see at night.”

Enhancing the senses to gain superior advantage—this language suggests weaponry. Such capacities could certainly have military applications, Ling acknowledged—“You can weaponize anything, right?”—before he dismissed the idea and returned to the party line: “No, actually, this has to do with increasing a human’s capability” in a way that he compared to military training and civilian education, and justified in economic terms.

“Let’s say I gave you a third arm,” and then a fourth arm—so, two additional hands, he said. “You would be more capable; you would do more things, right?” And if you could control four hands as seamlessly as you’re controlling your current two hands, he continued, “you would actually be doing double the amount of work that you would normally do. It’s as simple as that. You’re increasing your productivity to do whatever you want to do.” I started to picture his vision—working with four arms, four hands—and asked, “Where does it end?”

“It won’t ever end,” Ling said. “I mean, it will constantly get better and better—” His cellphone rang. He took the call, then resumed where he had left off: “What darpa does is we provide a fundamental tool so that other people can take those tools and do great things with them that we’re not even thinking about.”

Judging by what he said next, however, the number of things that darpa is thinking about far exceeds what it typically talks about in public. “If a brain can control a robot that looks like a hand,” Ling said, “why can’t it control a robot that looks like a snake? Why can’t that brain control a robot that looks like a big mass of Jell-O, able to get around corners and up and down and through things? I mean, somebody will find an application for that. They couldn’t do it now, because they can’t become that glob, right? But in my world, with their brain now having a direct interface with that glob, that glob is the embodiment of them. So now they’re basically the glob, and they can go do everything a glob can do.”

VI. Gold Rush

darpa’s developing capabilities still hover at or near a proof-of-concept stage. But that’s close enough to have drawn investment from some of the world’s richest corporations. In 1990, during the administration of President George H. W. Bush, darpa Director Craig I. Fields lost his job because, according to contemporary news accounts, he intentionally fostered business development with some Silicon Valley companies, and White House officials deemed that inappropriate. Since the administration of the second President Bush, however, such sensitivities have faded.

Over time, darpa has become something of a farm team for Silicon Valley. Regina Dugan, who was appointed darpa director by President Barack Obama, went on to head Google’s Advanced Technology and Projects group, and other former darpa officials went to work for her there. She then led R&D for the analogous group at Facebook, called Building 8. (She has since left Facebook.)

darpa’s neurotechnology research has been affected in recent years by corporate poaching. Doug Weber told me that some darpa researchers have been “scooped up” by companies including Verily, the life-sciences division of Alphabet (the parent company of Google), which, in partnership with the British pharmaceutical conglomerate GlaxoSmithKline, created a company called Galvani Bioelectronics, to bring neuro-modulation devices to market. Galvani calls its business “bioelectric medicine,” which conveys an aura of warmth and trustworthiness. Ted Berger, a University of Southern California biomedical engineer who collaborated with the Wake Forest researchers on their studies of memory transfer in rats, worked as the chief science officer at the neurotechnology company Kernel, which plans to build “advanced neural interfaces to treat disease and dysfunction, illuminate the mechanisms of intelligence, and extend cognition.” Elon Musk has courted darpa researchers to join his company Neuralink, which is said to be developing an interface known as “neural lace.” Facebook’s Building 8 is working on a neural interface too. In 2017, Regina Dugan said that 60 engineers were at work on a system with the goal of allowing users to type 100 words a minute “directly from your brain.” Geoff Ling is on Building 8’s advisory board.

Talking with Justin Sanchez, I speculated that if he realizes his ambitions, he could change daily life in even more fundamental and lasting ways than Facebook’s Mark Zuckerberg and Twitter’s Jack Dorsey have. Sanchez blushes easily, and he breaks eye contact when he is uncomfortable, but he did not look away when he heard his name mentioned in such company. Remembering a remark that he had once made about his hope for neurotechnology’s wide adoption, but with “appropriate checks to make sure that it’s done in the right way,” I asked him to talk about what the right way might look like. Did any member of Congress strike him as having good ideas about legal or regulatory structures that might shape an emerging neural-interface industry? He demurred (“darpa’s mission isn’t to define or even direct those things”) and suggested that, in reality, market forces would do more to shape the evolution of neurotechnology than laws or regulations or deliberate policy choices. What will happen, he said, is that scientists at universities will sell their discoveries or create start-ups. The marketplace will take it from there: “As they develop their companies, and as they develop their products, they’re going to be subject to convincing people that whatever they’re developing makes sense, that it helps people to be a better version of themselves. And that process—that day-to-day development—will ultimately guide where these technologies go. I mean, I think that’s the frank reality of how it ultimately will unfold.”

He seemed entirely untroubled by what may be the most troubling aspect of darpa’s work: not that it discovers what it discovers, but that the world has, so far, always been ready to buy it.

This article appears in the November 2018 print edition with the headline “The Pentagon Wants to Weaponize the Brain. What Could Go Wrong?”


Google’s DeepMind AI can accurately detect 50 types of eye disease just by looking at scans

Mustafa Suleyman 1831_preview (1)DeepMind cofounder Mustafa Suleyman.DeepMind
  • Google’s artificial intelligence company DeepMind has published „really significant“ research showing its algorithm can identify around 50 eye diseases by looking at retinal eye scans.
  • DeepMind said its AI was as good as expert clinicians, and that it could help prevent people from losing their sight.
  • DeepMind has been criticised for its practices around medical data, but cofounder Mustafa Suleyman said all the information in this research project was anonymised.
  • The company plans to hand the technology over for free to NHS hospitals for five years, provided it passes the next phase of research.

Google’s artificial intelligence company, DeepMind, has developed an AI which can successfully detect more than 50 types of eye disease just by looking at 3D retinal scans.

DeepMind published on Monday the results of joint research with Moorfields Eye Hospital, a renowned centre for treating eye conditions in London, in Nature Medicine.

The company said its AI was as accurate as expert clinicians when it came to detecting diseases, such as diabetic eye disease and macular degeneration. It could also recommend the best course of action for patients and suggest which needed urgent care.

OCT scanA technician examines an OCT scan.DeepMind

What is especially significant about the research, according to DeepMind cofounder Mustafa Suleyman, is that the AI has a level of „explainability“ that could boost doctors‘ trust in its recommendations.

„It’s possible for the clinician to interpret what the algorithm is thinking,“ he told Business Insider. „[They can] look at the underlying segmentation.“

In other words, the AI looks less like a mysterious black box that’s spitting out results. It labels pixels on the eye scan that corresponds to signs of a particular disease, Suleyman explained, and can calculate its confidence in its own findings with a percentage score. „That’s really significant,“ he said.

DeepMind's algorithm analysing an OCT eye scanDeepMind’s AI analysing an OCT scan.DeepMind

Suleyman described the findings as a „research breakthrough“ and said the next step was to prove the AI works in a clinical setting. That, he said, would take a number of years. Once DeepMind is in a position to deploy its AI across NHS hospitals in the UK, it will provide the service for free for five years.

Patients are at risk of losing their sight because doctors can’t look at their eye scans in time

British eye specialists have been warning for years that patients are at risk of losing their sight because the NHS is overstretched, and because the UK has an ageing population.

Part of the reason DeepMind and Moorfields took up the research project was because clinicians are „overwhelmed“ by the demand for eye scans, Suleyman said.

„If you have a sight-threatening disease, you want treatment as soon as possible,“ he explained. „And unlike in A&E, where a staff nurse will talk to you and make an evaluation of how serious your condition is, then use that evaluation to decide how quickly you are seen. When an [eye] scan is submitted, there isn’t a triage of your scan according to its severity.“

OCT scanA patient having an OCT scan.DeepMind

Putting eye scans through the AI could speed the entire process up.

„In the future, I could envisage a person going into their local high street optician, and have an OCT scan done and this algorithm would identify those patients with sight-threatening disease at the very early stage of the condition,“ said Dr Pearse Keane, consultant ophthalmologist at Moorfields Eye Hospital.

DeepMind’s AI was trained on a database of almost 15,000 eye scans, stripped of any identifying information. DeepMind worked with clinicians to label areas of disease, then ran those labelled images through its system. Suleyman said the two-and-a-half project required „huge investment“ from DeepMind and involved 25 staffers, as well as the researchers from Moorfields.

People are still worried about a Google-linked company having access to medical data

Google acquired DeepMind in 2014 for £400 million ($509 million), and the British AI company is probably most famous for AlphaGo, its algorithm that beat the world champion at the strategy game Go.

While DeepMind has remained UK-based and independent from Google, the relationship has attracted scrutiny. The main question is whether Google, a private US company, should have access to the sensitive medical data required for DeepMind’s health arm.

DeepMind was criticised in 2016 for failing to disclose its access to historical medical data during a project with Royal Free Hospital. Suleyman said the eye scans processed by DeepMind were „completely anonymised.“

„You can’t identify whose scans it was. We’re in quite a different regime, this is very much research, and we’re a number of years from being able to deploy in practice,“ he said.

Suleyman added: „How this has the potential to have transform the NHS is very clear. We’ve been very conscious that this will be a model that’s published, and available to others to implement.

„The labelled dataset is available to other researchers. So this is very much an open and collaborative relationship between equals that we’ve worked hard to foster. I’m proud of that work.“

Microsoft wants regulation of facial recognition technology to limit ‚abuse‘

Facial recognition put to the test
Facial recognition put to the test

Microsoft has helped innovate facial recognition software. Now it’s urging the US government to enact regulation to control the use of the technology.

In a blog post, Microsoft (MSFT)President Brad Smith said new laws are necessary given the technology’s „broad societal ramifications and potential for abuse.“

He urged lawmakers to form „a government initiative to regulate the proper use of facial recognition technology, informed first by a bipartisan and expert commission.“

Facial recognition — a computer’s ability to identify or verify people’s faces from a photo or through a camera — has been developing rapidly. Apple (AAPL), Google (GOOG), Amazon and Microsoft are among the big tech companies developing and selling such systems. The technology is being used across a range of industries, from private businesses like hotels and casinos, to social media and law enforcement.

Supporters say facial recognition software improves safety for companies and customers and can help police track police down criminals or find missing children. Civil rights groups warn it can infringe on privacy and allow for illegal surveillance and monitoring. There is also room for error, they argue, since the still-emerging technology can result in false identifications.

The accuracy of facial recognition technologies varies, with women and people of color being identified with less accuracy, according to MIT research.

„Facial recognition raises a critical question: what role do we want this type of technology to play in everyday society?“ Smith wrote on Friday.

Smith’s call for a regulatory framework to control the technology comes as tech companies face criticism over how they’ve handled and shared customer data, as well as their cooperation with government agencies.

Last month, Microsoft was scrutinized for its working relationship with US Immigration and Customs Enforcement. ICE had been enforcing the Trump administration’s „zero tolerance“ immigration policy that separated children from their parents when they crossed the US border illegally. The administration has since abandoned the policy.

Microsoft urges Trump administration to change its policy separating families at border

Microsoft wrote a blog post in January about ICE’s use of its cloud technology Azure, saying it could help it „accelerate facial recognition and identification.“

After questions arose about whether Microsoft’s technology had been used by ICE agents to carry out the controversial border separations, the company released a statement calling the policy „cruel“ and „abusive.“

In his post, Smith reiterated Microsoft’s opposition to the policy and said he had confirmed its contract with ICE does not include facial recognition technology.

Amazon(AMZN) has also come under fire from its own shareholders and civil rights groups over local police forces using its face identifying software Rekognition, which can identify up to 100 people in a single photo.

Some Amazon shareholders coauthored a letter pressuring Amazon to stop selling the technology to the government, saying it was aiding in mass surveillance and posed a threat to privacy rights.

Amazon asked to stop selling facial recognition technology to police

And Facebook (FB) is embroiled in a class-action lawsuit that alleges the social media giant used facial recognition on photos without user permission. Its facial recognition tool scans your photos and suggests you tag friends.

Neither Amazon nor Facebook immediately responded to a request for comment about Smith’s call for new regulations on face ID technology.

Smith said companies have a responsibility to police their own innovations, control how they are deployed and ensure that they are used in a „a manner consistent with broadly held societal values.“

„It may seem unusual for a company to ask for government regulation of its products, but there are many markets where thoughtful regulation contributes to a healthier dynamic for consumers and producers alike,“ he said.

The Evolution of AI

Photo credit: Peg Skorpinski


Artificial Intelligence — The Revolution Hasn’t Happened Yet

Artificial Intelligence (AI) is the mantra of the current era. The phrase is intoned by technologists, academicians, journalists and venture capitalists alike. As with many phrases that cross over from technical academic fields into general circulation, there is significant misunderstanding accompanying the use of the phrase. But this is not the classical case of the public not understanding the scientists — here the scientists are often as befuddled as the public. The idea that our era is somehow seeing the emergence of an intelligence in silicon that rivals our own entertains all of us — enthralling us and frightening us in equal measure. And, unfortunately, it distracts us.

There is a different narrative that one can tell about the current era. Consider the following story, which involves humans, computers, data and life-or-death decisions, but where the focus is something other than intelligence-in-silicon fantasies. When my spouse was pregnant 14 years ago, we had an ultrasound. There was a geneticist in the room, and she pointed out some white spots around the heart of the fetus. “Those are markers for Down syndrome,” she noted, “and your risk has now gone up to 1 in 20.” She further let us know that we could learn whether the fetus in fact had the genetic modification underlying Down syndrome via an amniocentesis. But amniocentesis was risky — the risk of killing the fetus during the procedure was roughly 1 in 300. Being a statistician, I determined to find out where these numbers were coming from. To cut a long story short, I discovered that a statistical analysis had been done a decade previously in the UK, where these white spots, which reflect calcium buildup, were indeed established as a predictor of Down syndrome. But I also noticed that the imaging machine used in our test had a few hundred more pixels per square inch than the machine used in the UK study. I went back to tell the geneticist that I believed that the white spots were likely false positives — that they were literally “white noise.” She said “Ah, that explains why we started seeing an uptick in Down syndrome diagnoses a few years ago; it’s when the new machine arrived.”

We didn’t do the amniocentesis, and a healthy girl was born a few months later. But the episode troubled me, particularly after a back-of-the-envelope calculation convinced me that many thousands of people had gotten that diagnosis that same day worldwide, that many of them had opted for amniocentesis, and that a number of babies had died needlessly. And this happened day after day until it somehow got fixed. The problem that this episode revealed wasn’t about my individual medical care; it was about a medical system that measured variables and outcomes in various places and times, conducted statistical analyses, and made use of the results in other places and times. The problem had to do not just with data analysis per se, but with what database researchers call “provenance” — broadly, where did data arise, what inferences were drawn from the data, and how relevant are those inferences to the present situation? While a trained human might be able to work all of this out on a case-by-case basis, the issue was that of designing a planetary-scale medical system that could do this without the need for such detailed human oversight.

I’m also a computer scientist, and it occurred to me that the principles needed to build planetary-scale inference-and-decision-making systems of this kind, blending computer science with statistics, and taking into account human utilities, were nowhere to be found in my education. And it occurred to me that the development of such principles — which will be needed not only in the medical domain but also in domains such as commerce, transportation and education — were at least as important as those of building AI systems that can dazzle us with their game-playing or sensorimotor skills.

Whether or not we come to understand “intelligence” any time soon, we do have a major challenge on our hands in bringing together computers and humans in ways that enhance human life. While this challenge is viewed by some as subservient to the creation of “artificial intelligence,” it can also be viewed more prosaically — but with no less reverence — as the creation of a new branch of engineering. Much like civil engineering and chemical engineering in decades past, this new discipline aims to corral the power of a few key ideas, bringing new resources and capabilities to people, and doing so safely. Whereas civil engineering and chemical engineering were built on physics and chemistry, this new engineering discipline will be built on ideas that the preceding century gave substance to — ideas such as “information,” “algorithm,” “data,” “uncertainty,” “computing,” “inference,” and “optimization.” Moreover, since much of the focus of the new discipline will be on data from and about humans, its development will require perspectives from the social sciences and humanities.

While the building blocks have begun to emerge, the principles for putting these blocks together have not yet emerged, and so the blocks are currently being put together in ad-hoc ways.

Thus, just as humans built buildings and bridges before there was civil engineering, humans are proceeding with the building of societal-scale, inference-and-decision-making systems that involve machines, humans and the environment. Just as early buildings and bridges sometimes fell to the ground — in unforeseen ways and with tragic consequences — many of our early societal-scale inference-and-decision-making systems are already exposing serious conceptual flaws.

And, unfortunately, we are not very good at anticipating what the next emerging serious flaw will be. What we’re missing is an engineering discipline with its principles of analysis and design.

The current public dialog about these issues too often uses “AI” as an intellectual wildcard, one that makes it difficult to reason about the scope and consequences of emerging technology. Let us begin by considering more carefully what “AI” has been used to refer to, both recently and historically.

Most of what is being called “AI” today, particularly in the public sphere, is what has been called “Machine Learning” (ML) for the past several decades. ML is an algorithmic field that blends ideas from statistics, computer science and many other disciplines (see below) to design algorithms that process data, make predictions and help make decisions. In terms of impact on the real world, ML is the real thing, and not just recently. Indeed, that ML would grow into massive industrial relevance was already clear in the early 1990s, and by the turn of the century forward-looking companies such as Amazon were already using ML throughout their business, solving mission-critical back-end problems in fraud detection and supply-chain prediction, and building innovative consumer-facing services such as recommendation systems. As datasets and computing resources grew rapidly over the ensuing two decades, it became clear that ML would soon power not only Amazon but essentially any company in which decisions could be tied to large-scale data. New business models would emerge. The phrase “Data Science” began to be used to refer to this phenomenon, reflecting the need of ML algorithms experts to partner with database and distributed-systems experts to build scalable, robust ML systems, and reflecting the larger social and environmental scope of the resulting systems.

This confluence of ideas and technology trends has been rebranded as “AI” over the past few years. This rebranding is worthy of some scrutiny.

Historically, the phrase “AI” was coined in the late 1950’s to refer to the heady aspiration of realizing in software and hardware an entity possessing human-level intelligence. We will use the phrase “human-imitative AI” to refer to this aspiration, emphasizing the notion that the artificially intelligent entity should seem to be one of us, if not physically at least mentally (whatever that might mean). This was largely an academic enterprise. While related academic fields such as operations research, statistics, pattern recognition, information theory and control theory already existed, and were often inspired by human intelligence (and animal intelligence), these fields were arguably focused on “low-level” signals and decisions. The ability of, say, a squirrel to perceive the three-dimensional structure of the forest it lives in, and to leap among its branches, was inspirational to these fields. “AI” was meant to focus on something different — the “high-level” or “cognitive” capability of humans to “reason” and to “think.” Sixty years later, however, high-level reasoning and thought remain elusive. The developments which are now being called “AI” arose mostly in the engineering fields associated with low-level pattern recognition and movement control, and in the field of statistics — the discipline focused on finding patterns in data and on making well-founded predictions, tests of hypotheses and decisions.

Indeed, the famous “backpropagation” algorithm that was rediscovered by David Rumelhart in the early 1980s, and which is now viewed as being at the core of the so-called “AI revolution,” first arose in the field of control theory in the 1950s and 1960s. One of its early applications was to optimize the thrusts of the Apollo spaceships as they headed towards the moon.

Since the 1960s much progress has been made, but it has arguably not come about from the pursuit of human-imitative AI. Rather, as in the case of the Apollo spaceships, these ideas have often been hidden behind the scenes, and have been the handiwork of researchers focused on specific engineering challenges. Although not visible to the general public, research and systems-building in areas such as document retrieval, text classification, fraud detection, recommendation systems, personalized search, social network analysis, planning, diagnostics and A/B testing have been a major success — these are the advances that have powered companies such as Google, Netflix, Facebook and Amazon.

One could simply agree to refer to all of this as “AI,” and indeed that is what appears to have happened. Such labeling may come as a surprise to optimization or statistics researchers, who wake up to find themselves suddenly referred to as “AI researchers.” But labeling of researchers aside, the bigger problem is that the use of this single, ill-defined acronym prevents a clear understanding of the range of intellectual and commercial issues at play.

The past two decades have seen major progress — in industry and academia — in a complementary aspiration to human-imitative AI that is often referred to as “Intelligence Augmentation” (IA). Here computation and data are used to create services that augment human intelligence and creativity. A search engine can be viewed as an example of IA (it augments human memory and factual knowledge), as can natural language translation (it augments the ability of a human to communicate). Computing-based generation of sounds and images serves as a palette and creativity enhancer for artists. While services of this kind could conceivably involve high-level reasoning and thought, currently they don’t — they mostly perform various kinds of string-matching and numerical operations that capture patterns that humans can make use of.

Hoping that the reader will tolerate one last acronym, let us conceive broadly of a discipline of “Intelligent Infrastructure” (II), whereby a web of computation, data and physical entities exists that makes human environments more supportive, interesting and safe. Such infrastructure is beginning to make its appearance in domains such as transportation, medicine, commerce and finance, with vast implications for individual humans and societies. This emergence sometimes arises in conversations about an “Internet of Things,” but that effort generally refers to the mere problem of getting “things” onto the Internet — not to the far grander set of challenges associated with these “things” capable of analyzing those data streams to discover facts about the world, and interacting with humans and other “things” at a far higher level of abstraction than mere bits.

For example, returning to my personal anecdote, we might imagine living our lives in a “societal-scale medical system” that sets up data flows, and data-analysis flows, between doctors and devices positioned in and around human bodies, thereby able to aid human intelligence in making diagnoses and providing care. The system would incorporate information from cells in the body, DNA, blood tests, environment, population genetics and the vast scientific literature on drugs and treatments. It would not just focus on a single patient and a doctor, but on relationships among all humans — just as current medical testing allows experiments done on one set of humans (or animals) to be brought to bear in the care of other humans. It would help maintain notions of relevance, provenance and reliability, in the way that the current banking system focuses on such challenges in the domain of finance and payment. And, while one can foresee many problems arising in such a system — involving privacy issues, liability issues, security issues, etc — these problems should properly be viewed as challenges, not show-stoppers.

We now come to a critical issue: Is working on classical human-imitative AI the best or only way to focus on these larger challenges? Some of the most heralded recent success stories of ML have in fact been in areas associated with human-imitative AI — areas such as computer vision, speech recognition, game-playing and robotics. So perhaps we should simply await further progress in domains such as these. There are two points to make here. First, although one would not know it from reading the newspapers, success in human-imitative AI has in fact been limited — we are very far from realizing human-imitative AI aspirations. Unfortunately the thrill (and fear) of making even limited progress on human-imitative AI gives rise to levels of over-exuberance and media attention that is not present in other areas of engineering.

Second, and more importantly, success in these domains is neither sufficient nor necessary to solve important IA and II problems. On the sufficiency side, consider self-driving cars. For such technology to be realized, a range of engineering problems will need to be solved that may have little relationship to human competencies (or human lack-of-competencies). The overall transportation system (an II system) will likely more closely resemble the current air-traffic control system than the current collection of loosely-coupled, forward-facing, inattentive human drivers. It will be vastly more complex than the current air-traffic control system, specifically in its use of massive amounts of data and adaptive statistical modeling to inform fine-grained decisions. It is those challenges that need to be in the forefront, and in such an effort a focus on human-imitative AI may be a distraction.

As for the necessity argument, it is sometimes argued that the human-imitative AI aspiration subsumes IA and II aspirations, because a human-imitative AI system would not only be able to solve the classical problems of AI (as embodied, e.g., in the Turing test), but it would also be our best bet for solving IA and II problems. Such an argument has little historical precedent. Did civil engineering develop by envisaging the creation of an artificial carpenter or bricklayer? Should chemical engineering have been framed in terms of creating an artificial chemist? Even more polemically: if our goal was to build chemical factories, should we have first created an artificial chemist who would have then worked out how to build a chemical factory?

A related argument is that human intelligence is the only kind of intelligence that we know, and that we should aim to mimic it as a first step. But humans are in fact not very good at some kinds of reasoning — we have our lapses, biases and limitations. Moreover, critically, we did not evolve to perform the kinds of large-scale decision-making that modern II systems must face, nor to cope with the kinds of uncertainty that arise in II contexts. One could argue
that an AI system would not only imitate human intelligence, but also “correct” it, and would also scale to arbitrarily large problems. But we are now in the realm of science fiction — such speculative arguments, while entertaining in the setting of fiction, should not be our principal strategy going forward in the face of the critical IA and II problems that are beginning to emerge. We need to solve IA and II problems on their own merits, not as a mere corollary to a human-imitative AI agenda.

It is not hard to pinpoint algorithmic and infrastructure challenges in II systems that are not central themes in human-imitative AI research. II systems require the ability to manage distributed repositories of knowledge that are rapidly changing and are likely to be globally incoherent. Such systems must cope with cloud-edge interactions in making timely, distributed decisions and they must deal with long-tail phenomena whereby there is lots of data on some individuals and little data on most individuals. They must address the difficulties of sharing data across administrative and competitive boundaries. Finally, and of particular importance, II systems must bring economic ideas such as incentives and pricing into the realm of the statistical and computational infrastructures that link humans to each other and to valued goods. Such II systems can be viewed as not merely providing a service, but as creating markets. There are domains such as music, literature and journalism that are crying out for the emergence of such markets, where data analysis links producers and consumers. And this must all be done within the context of evolving societal, ethical and legal norms.

Of course, classical human-imitative AI problems remain of great interest as well. However, the current focus on doing AI research via the gathering of data, the deployment of “deep learning” infrastructure, and the demonstration of systems that mimic certain narrowly-defined human skills — with little in the way of emerging explanatory principles — tends to deflect attention from major open problems in classical AI. These problems include the need to bring meaning and reasoning into systems that perform natural language processing, the need to infer and represent causality, the need to develop computationally-tractable representations of uncertainty and the need to develop systems that formulate and pursue long-term goals. These are classical goals in human-imitative AI, but in the current hubbub over the “AI revolution,” it is easy to forget that they are not yet solved.

IA will also remain quite essential, because for the foreseeable future, computers will not be able to match humans in their ability to reason abstractly about real-world situations. We will need well-thought-out interactions of humans and computers to solve our most pressing problems. And we will want computers to trigger new levels of human creativity, not replace human creativity (whatever that might mean).

It was John McCarthy (while a professor at Dartmouth, and soon to take a
position at MIT) who coined the term “AI,” apparently to distinguish his
budding research agenda from that of Norbert Wiener (then an older professor at MIT). Wiener had coined “cybernetics” to refer to his own vision of intelligent systems — a vision that was closely tied to operations research, statistics, pattern recognition, information theory and control theory. McCarthy, on the other hand, emphasized the ties to logic. In an interesting reversal, it is Wiener’s intellectual agenda that has come to dominate in the current era, under the banner of McCarthy’s terminology. (This state of affairs is surely, however, only temporary; the pendulum swings more in AI than
in most fields.)

But we need to move beyond the particular historical perspectives of McCarthy and Wiener.

We need to realize that the current public dialog on AI — which focuses on a narrow subset of industry and a narrow subset of academia — risks blinding us to the challenges and opportunities that are presented by the full scope of AI, IA and II.

This scope is less about the realization of science-fiction dreams or nightmares of super-human machines, and more about the need for humans to understand and shape technology as it becomes ever more present and influential in their daily lives. Moreover, in this understanding and shaping there is a need for a diverse set of voices from all walks of life, not merely a dialog among the technologically attuned. Focusing narrowly on human-imitative AI prevents an appropriately wide range of voices from being heard.

While industry will continue to drive many developments, academia will also continue to play an essential role, not only in providing some of the most innovative technical ideas, but also in bringing researchers from the computational and statistical disciplines together with researchers from other
disciplines whose contributions and perspectives are sorely needed — notably
the social sciences, the cognitive sciences and the humanities.

On the other hand, while the humanities and the sciences are essential as we go forward, we should also not pretend that we are talking about something other than an engineering effort of unprecedented scale and scope — society is aiming to build new kinds of artifacts. These artifacts should be built to work as claimed. We do not want to build systems that help us with medical treatments, transportation options and commercial opportunities to find out after the fact that these systems don’t really work — that they make errors that take their toll in terms of human lives and happiness. In this regard, as I have emphasized, there is an engineering discipline yet to emerge for the data-focused and learning-focused fields. As exciting as these latter fields appear to be, they cannot yet be viewed as constituting an engineering discipline.

Moreover, we should embrace the fact that what we are witnessing is the creation of a new branch of engineering. The term “engineering” is often
invoked in a narrow sense — in academia and beyond — with overtones of cold, affectless machinery, and negative connotations of loss of control by humans. But an engineering discipline can be what we want it to be.

In the current era, we have a real opportunity to conceive of something historically new — a human-centric engineering discipline.

I will resist giving this emerging discipline a name, but if the acronym “AI” continues to be used as placeholder nomenclature going forward, let’s be aware of the very real limitations of this placeholder. Let’s broaden our scope, tone down the hype and recognize the serious challenges ahead.

Michael I. Jordan


Most dangerous attack techniques, and what’s coming next 2018

RSA Conference 2018

Experts from SANS presented the five most dangerous new cyber attack techniques in their annual RSA Conference 2018 keynote session in San Francisco, and shared their views on how they work, how they can be stopped or at least slowed, and how businesses and consumers can prepare.

dangerous attack techniques

The five threats outlined are:

1. Repositories and cloud storage data leakage
2. Big Data analytics, de-anonymization, and correlation
3. Attackers monetize compromised systems using crypto coin miners
4. Recognition of hardware flaws
5. More malware and attacks disrupting ICS and utilities instead of seeking profit.

Repositories and cloud storage data leakage

Ed Skoudis, lead for the SANS Penetration Testing Curriculum, talked about the data leakage threats facing us from the increased use of repositories and cloud storage:

“Software today is built in a very different way than it was 10 or even 5 years ago, with vast online code repositories for collaboration and cloud data storage hosting mission-critical applications. However, attackers are increasingly targeting these kinds of repositories and cloud storage infrastructures, looking for passwords, crypto keys, access tokens, and terabytes of sensitive data.”

He continued: “Defenders need to focus on data inventories, appointing a data curator for their organization and educating system architects and developers about how to secure data assets in the cloud. Additionally, the big cloud companies have each launched an AI service to help classify and defend data in their infrastructures. And finally, a variety of free tools are available that can help prevent and detect leakage of secrets through code repositories.”

Big Data analytics, de-anonymization, and correlation

Skoudis went on to talk about the threat of Big Data Analytics and how attackers are using data from several sources to de-anonymise users:

“In the past, we battled attackers who were trying to get access to our machines to steal data for criminal use. Now the battle is shifting from hacking machines to hacking data — gathering data from disparate sources and fusing it together to de-anonymise users, find business weaknesses and opportunities, or otherwise undermine an organisation’s mission. We still need to prevent attackers from gaining shell on targets to steal data. However, defenders also need to start analysing risks associated with how their seemingly innocuous data can be combined with data from other sources to introduce business risk, all while carefully considering the privacy implications of their data and its potential to tarnish a brand or invite regulatory scrutiny.”

Attackers monetize compromised systems using crypto coin miners

Johannes Ullrich, is Dean of Research, SANS Institute and Director of SANS Internet Storm Center. He has been looking at the increasing use of crypto coin miners by cyber criminals:

“Last year, we talked about how ransomware was used to sell data back to its owner and crypto-currencies were the tool of choice to pay the ransom. More recently, we have found that attackers are no longer bothering with data. Due to the flood of stolen data offered for sale, the value of most commonly stolen data like credit card numbers of PII has dropped significantly. Attackers are instead installing crypto coin miners. These attacks are more stealthy and less likely to be discovered and attackers can earn tens of thousands of dollars a month from crypto coin miners. Defenders therefore need to learn to detect these coin miners and to identify the vulnerabilities that have been exploited in order to install them.”

Recognition of hardware flaws

Ullrich then went on to say that software developers often assume that hardware is flawless and that this is a dangerous assumption. He explains why and what needs to be done:

“Hardware is no less complex then software and mistakes have been made in developing hardware just as they are made by software developers. Patching hardware is a lot more difficult and often not possible without replacing entire systems or suffering significant performance penalties. Developers therefore need to learn to create software without relying on hardware to mitigate any security issues. Similar to the way in which software uses encryption on untrusted networks, software needs to authenticate and encrypt data within the system. Some emerging homomorphic encryption algorithms may allow developers to operate on encrypted data without having to decrypt it first.”

most dangerous attack techniques

More malware and attacks disrupting ICS and utilities instead of seeking profit

Finally, Head of R&D, SANS Institute, James Lyne, discussed the growing trend in malware and attacks that aren’t profit centred as we have largely seen in the past, but instead are focused on disrupting Industrial Control Systems (ICS) and utilities:

“Day to day the grand majority of malicious code has undeniably been focused on fraud and profit. Yet, with the relentless deployment of technology in our societies, the opportunity for political or even military influence only grows. And rare publicly visible attacks like Triton/TriSYS show the capability and intent of those who seek to compromise some of the highest risk components of industrial environments, i.e. the safety systems which have historically prevented critical security and safety meltdowns.”

He continued: “ICS systems are relatively immature and easy to exploit in comparison to the mainstream computing world. Many ICS systems lack the mitigations of modern operating systems and applications. The reliance on obscurity or isolation (both increasingly untrue) do not position them well to withstand a heightened focus on them, and we need to address this as an industry. More worrying is that attackers have demonstrated they have the inclination and resources to diversify their attacks, targeting the sensors that are used to provide data to the industrial controllers themselves. The next few years are likely to see some painful lessons being learned as this attack domain grows, since the mitigations are inconsistent and quite embryonic.”


Machine Learning – Basics – Einsatzgebiete – Technik

Machine Learning, Deep Learning, Cognitive Computing – Technologien der Künstlichen Intelligenz verbreiten sich rasant. Hintergrund ist, dass heute die Rechen- und Speicherkapazitäten zur Verfügung stehen, die KI-Szenarien möglich machen. Ein Überblick.
  • Machine Learning hilft, Muster in großen Datenbeständen zu erkennen und daraus Erkenntnisse zu gewinnen
  • Die Einsatzszenarien reichen von der Spamanalyse über Stauprognosen bis hin zur medizinischen Diagnostik
  • Technische Grundlage ist eine Cloud-basierte Digital Infrastructure Platform,3330413,3330418,3330420

Künstliche Intelligenz und Machine Learning (ML) sind keine neuen Technologien, doch im praktischen Einsatz spielen sie erst jetzt eine wichtige Rolle. Woran liegt das? Wichtigste Voraussetzung für lernende Systeme und entsprechende Algorithmen sind ausreichende Rechenkapazitäten und der Zugriff auf riesige Datenmengen – egal ob es sich um Kunden-, Log- oder Sensordaten handelt. Sie sind für das Training der Algorithmen und die Modellbildung unverzichtbar – und sie stehen mit Public- und Private-Cloud-Infrastrukturen zur Verfügung.

Bildanalyse und -erkennung ist das wichtigste Machine-Learning-Thema, doch die Spracherkennung und -verarbeitung ist schwer im Kommen.
Bildanalyse und -erkennung ist das wichtigste Machine-Learning-Thema, doch die Spracherkennung und -verarbeitung ist schwer im Kommen.
Foto: Crisp Research, Kassel


Die Analysten von Crisp Research sind im Rahmen einer umfassenden Studie gemeinsam mit The unbelievable Machine Company und Hewlett-Packard Enterprise (HPE) der Frage nachgegangen, welche Rolle Machine Learning heute und in Zukunft im Unternehmenseinsatz spielen wird. Dabei zeigt sich, dass deutsche Unternehmen hier schon recht weit fortgeschritten sind. Bereits ein Fünftel setzt ML-Technologien aktiv ein, 64 Prozent beschäftigen sich intensiv mit dem Thema und vier von fünf Befragten sagen sogar, ML werde irgendwann eine der Kerntechnologien des vollständig digitalisierten Unternehmens sein.

Muster erkennen und Vorhersagen treffen

ML-Algorithmen helfen den Menschen, Muster in vorhandenen Datenbeständen zu erkennen, Vorhersagen zu treffen oder Daten zu klassifizieren. Mit mathematischen Modellen können neue Erkenntnisse auf Grundlage dieser Muster gewonnen werden. Das gilt für viele Lebens- und Geschäftsbereiche. Oftmals profitieren Internet-Nutzer längst davon, ohne über die Technologie im Hintergrund nachzudenken.

Das Spektrum der Anwendungen reicht von Musik- und Filmempfehlungen im privaten Umfeld bis hin zur Verbesserung von Marketing-Kampagnen, Kundenservices oder auch Logistikrouten im geschäftlichen Bereich. Dafür steht ein breites Spektrum an ML-Verfahren zur Verfügung, darunter Lineare Regression, Instanzenbasiertes Lernen, Entscheidungs-Baum-Algorithmen, Bayesche Statistik, Clusteranalyse, Neuronale Netzwerke, Deep Learning und Verfahren zur Dimensionsreduktion.

Die Anwendungsbereiche sind vielfältig und teilweise bekannt. Man denke etwa an Spam-Erkennung, die Personalisierung von Inhalten, das Klassifizieren von Dokumenten, Sentiment-Analysen, Prognosen der Kundenabwanderung, E-Mail-Klassifizierung, Analyse von Upselling-Möglichkeiten, Stauprognosen, Genomanalysen, medizinische Diagnostik, Chatbots und vieles mehr. Für nahezu alle Branchen und Unternehmenstypen ergeben sich also Gelegenheiten.

Moderne IT-Plattformen unterstützen KI

Machine Learning ist laut Crisp Research idealerweise Bestandteil einer modernen, skalierungsfähigen und flexiblen IT-Infrastruktur – einer „Digital Infrastructure Platform“. Diese zeichnet sich durch Elastizität, Automatisierung, eine API-basierte Architektur und Agilität aus. Eine solche Plattform ist in der Regel Cloud-basiert aufgesetzt und dient als Grundlage für die Entwicklung und den Betrieb neuer digitaler Anwendungen und Prozesse. Sie bietet eine offene Architektur, Programmierschnittstellen (APIs), um externe Services zu integrieren, die Unterstützung von DevOps-Konzepten sowie moderne Methoden für kurze Release- und Innovationszyklen.

Die Verarbeitung und Analyse großer Datenmengen ist eine Kernaufgabe einer solchen Digital Infrastructure Platform. Deshalb müssen die IT-Verantwortlichen Sorge tragen, dass ihre IT mit unterschiedlichen Verfahren der Künstlichen Intelligenz umgehen kann. Server-, Storage- und Netzwerk-Infrastrukturen müssen auf neue ML-basierte Workloads ausgelegt sein. Auch das Daten-Management muss vorbereitet sein, damit ML-as-a-Service-Angebote in der Cloud genutzt werden können.

Im Kontext von ML haben sich in den vergangenen Monaten auch alternative Hardwarekomponenten durchgesetzt, etwa GPU-basierte Cluster von Nvidia, Googles Tensor Processing Unit (TPU) oder IBMs TrueNorth-Prozessor. Unternehmen müssen sich entscheiden, ob sie hier selbst investieren oder die Angebote entsprechender Cloud-Provider nutzen wollen.

Einer der großen Anwendungsbereiche für ML ist die Spracherkennung und -verarbeitung. Amazons Alexa zieht gerade in die Haushalte ein, Microsoft, Google, Facebook und IBM haben hier einen Großteil ihrer Forschungs- und Entwicklungsgelder investiert sowie spezialisierte Firmen zugekauft. Es lässt sich absehen, dass natürlichsprachige Kommunikation an der Kundenschnittstelle selbstverständlicher wird. Auch die Bedienung von digitalen Produkten und Enterprise-IT-Lösungen wird via Sprachbefehl möglich sein. Das hat sowohl Auswirkungen auf das Customer-Frontend als auch auf das IT-Backend.

Niedrige Einstiegshürden in Machine Learning

Da die großen Cloud-Anbieter ML-Services und -Produkte in ihr Leistungsportfolio aufgenommen haben, ist es für Anwender relativ einfach, einen Einstieg zu finden. Amazon Machine Learning, Microsoft Azure Machine Learning, IBM Bluemix und Google Machine Learning erlauben einen kostengünstigen Zugang zu entsprechenden Diensten über die Public Cloud. Anwender brauchen also keinen eigenen Supercomputer, kein Team von Statistikexperten und kein dediziertes Infrastruktur-Management mehr. Mit ein paar Kommandos über die APIs der großen Public-Cloud-Provider können sie loslegen.

Anwender brauchen vor allem Hilfe bei der Datenexploration.
Anwender brauchen vor allem Hilfe bei der Datenexploration.
Foto: Crisp Research, Kassel


Sie finden dort unterschiedliche Machine-Learning-Verfahren sowie Dienste und Tools wie etwa grafische Programmiermodelle und Storage-Dienste vor. Je mehr sie sich darauf einlassen, desto größer wird allerdings das Risiko eines Vendor-Lock-ins. Deshalb sollten sich Anwender vor dem Start Gedanken über ihre Strategie machen. IT-Dienstleister und Managed-Service-Provider können ebenso ML-Systeme und Infrastrukturen bereitstellen und betreiben, so dass Unabhängigkeit von den Public-Cloud-Providern und ihren SLAs ebenso möglich ist.

Verschiedene Spielarten der KI

Machine Learning, Deep Learning, Cognitive Computing – derzeit kursieren eine Reihe von KI-Begriffen, deren Abgrenzung voneinander nicht ganz einfach ist. Crisp Research wählt dafür die Dimensionen „Clarity of Purpose“ (Orientierung am Einsatzweck) und „Degree of Autonomy“ (Grad der Autonomie). ML-Systeme sind derzeit größtenteils auf Einsatzzwecke hin entwickelt und trainiert. Sie erkennen beispielsweise im Fertigungsprozess fehlerhafte Produkte im Rahmen einer Qualitätskontrolle. Ihre Aufgabe ist klar umrissen, es gibt keine Autonomie.

Deep-Learning-Systeme hingegen sind in der Lage, mittels Neuronaler Netze eigenständig zu lernen. Simulierte Neuronen werden in vielen Schichten übereinander modelliert und angeordnet. Jede Ebene des Netzwerks erfüllt dabei eigenständig bestimmte Aufgaben, etwa das Erkennen von Kanten. Diese Information wird eigenständig an die nächste Ebene weitergegeben und fließt dort in die Verarbeitung ein. Im Zusammenspiel mit großen Mengen an Trainingsdaten lernen solche Netzwerke, bestimmte Aufgaben zu erledigen – etwa das Identifizieren von Krebszellen in medizinischen Bildern.

Deep-Learning-Systeme arbeiten autonomer

Deep-Learning-Systeme arbeiten also deutlich autonomer als ML-Systeme, da die Neuronalen Netzwerke darauf trainiert werden, selbständig zu lernen und Entscheidungen zu treffen, die von außen nicht unbedingt nachvollziehbar sind.

Als dritte Spielart der KI gilt das Cognitive Computing, das insbesondere von IBM mit seiner Watson-Technologie propagiert wird. Solche Systeme zeichnen sich dadurch aus, dass sie in einer Assistenzfunktion oder gar als Ersatz des Menschen Aufgaben übernehmen und Entscheidungen treffen und dabei mit Ambiguität und Unschärfe umgehen können. Als Beispiele können das Schadensfall-Management in einer Versicherung dienen, eine Service-Hotline oder die Diagnostik im Krankenhaus.

Auch wenn hier bereits ein hohes Maß an Autonomie erreicht werden kann, ist der Weg zu echter Künstlicher Intelligenz mit autonomen kognitiven Fähigkeiten noch weit. Die Wissenschaft beschäftigt sich aber intensiv damit und streitet darüber, ob und wann dieses Ziel erreicht werden kann. Derweil sind Unternehmen gut beraten, sich mit den machbaren Use Cases zu beschäftigen, von denen es bereits eine Menge gibt.

Im Zuge des Digitalisierungstrends kommt in vielen Unternehmen Analytics auf die Tagesordnung – und damit auch Machine Learning und Deep Learning. Jetzt geht es darum, den Datenschatz zu heben.
  • Viele Unternehmen haben Data Lakes mit strukturierten und unstrukturierten Daten aufgebaut. Jetzt gilt es, etwas daraus zu machen
  • Einsatzgebiete für Machine Learning sind etwa Prozessverbesserungen sowie eine bessere Kundenansprache und ein möglichst effizienter Support
  • In vielen Branchen ist der Abstand zwischen Vorreitern und Nachzüglern riesig

Die Phantasien und Visionen rund um die digitale Zukunft kennen derzeit keine Grenzen. Vollautomatisierte Produktionsstraßen, autonome Verkehrssysteme, intelligente digitale Assistenten – es vergeht kaum ein Tag, an dem nicht neue Szenarien diskutiert werden. Dadurch fühlen sich viele Firmen unter Druck gesetzt. Sie arbeiten am „digitalen Unternehmen“ und entdecken ihre Daten als Grundlagen für neue Geschäftsmodelle und Services. So gewinnt Analytics an Bedeutung – und mit der Analytics-Strategie kommen KI und Machine Learning (ML) auf die Tagesordnung.

Aus diesen Gründen beschäftigen sich Anwender mit Machine Learning.
Aus diesen Gründen beschäftigen sich Anwender mit Machine Learning.
Foto: Crisp Research


IT- und Digitalisierungsentscheider vermuten ein enormes Potenzial hinter dem Thema Machine Learning. Eine Umfrage, die das Analystenhaus Crisp Research unterstützt von The unbelievable Machine Company und Hewlett-Packard Enterprise (HPE) auf den Weg gebracht hat, zeigt, dass nur drei Prozent der knapp 250 Befragten ML für einen Marketing-Hype halten. Ein Drittel bezeichnet ML-Verfahren in begrenzten Einsatzbereichen als sinnvoll, sogar 43 Prozent sind überzeugt davon, dass ML ein wichtiger Aspekt künftiger Big-Data- und Analytics-Strategien wird.

Wie die Initiatoren der Studie feststellen, ist das kein überraschendes Ergebnis. Die meisten Unternehmen haben im großen Stil in Big-Data-Infrastrukturen und eigene Data Lakes investiert, um ihre Unternehmensdaten zusammenzuführen und auswertbar zu machen. ML ermöglicht einen hohen Automationsgrad in der Datenanalyse und hilft somit, den verborgenen Schatz zu heben. Daten gelten als großes Asset, doch den Beweis dafür haben viele Firmen noch nicht gebracht. Technologien und Use Cases rund um Machine Learning versprechen Abhilfe.

Immenses Innovationspotenzial

Immerhin 16 Prozent der befragten sehen ML sogar als neue „Kerntechnologie eines vollständig digitalen Unternehmens“. Das Innovations- und Gestaltungspotenzial scheint also immens, wenngleich viele Probleme rund um Datenqualität, Governance, API-Management, Infrastruktur und vor allem Personal den Trend noch bremsen.

Rund 34 Prozent der Befragten beschäftigen sich mit ML, weil sie ihre internen Prozesse in der Produktion, Logistik oder im Qualitätsmanagement verbessern wollen. Sie erheben beispielsweise Daten im Produktionsablauf, um ihre Fertigung optimieren zu können. Fast ebenso viele wollen Initiativen rund um die Customer Experience vorantreiben – etwa in E-Commerce, Marketing oder im Bereich der Portale und Apps. Sie versprechen sich davon beispielsweise eine personalisierte Kundenansprache, um Produkte oder Dienste zielgerichteter an den Konsumenten bringen zu können. Mit 19 Prozent ist die Gruppe derer, die Wartungs- und Supportleistungen optimieren wollen (Predictive Maintenance), etwas kleiner. Hinzu kommen Betriebe, die sich grundsätzlich mit neuen Technologien beschäftigen (28 Prozent) oder durch Berater und Analysten auf das Thema aufmerksam geworden sind (27 Prozent).

Elementar für selbstfahrende Autos

Das Nutzungsverhalten von ML ist nicht nur zwischen, sondern auch innerhalb der Branchen sehr unterschiedlich ausgeprägt. In der Automobilbranche etwa gibt es große Abstände zwischen den Vorreitern und den Nachzüglern. Für die Entwicklung und Produktion selbstfahrender Autos sind Bild- und Videoanalyse in Echtzeit sowie statistische Verfahren und mathematische Modelle aus Machine Learning und Deep Learning weit verbreitet. Einige Verfahren werden auch dazu verwendet, Fabrikationsfehler in der Fertigung zu erkennen.

Der Anteil der Innovatoren, die ML bereits in weiten Teilen einsetzen, ist in der Automobilbranche mit rund 20 Prozent am größten. Demgegenüber stehen allerdings 60 Prozent, die sich zwar mit ML beschäftigen, aber noch in der Evaluierungs- und Planungsphase stecken. So zeigt sich, dass in der Autobranche einige Leuchttürme das Bild prägen, von einer flächendeckenden Adaption aber nicht die Rede sein kann.

Status der Branchen bei der Einführung von Machine-Learning-Technologien
Status der Branchen bei der Einführung von Machine-Learning-Technologien
Foto: Crisp Research


Auch die Maschinen- und Anlagenbauer stecken noch zur Hälfte (53 Prozent) in der Evaluierungs- und Planungsphase. Ein knappe Drittel nutzt ML in ausgewählten Anwendungsbereichen produktiv und 18 Prozent bauen derzeit Prototypen. Weiter sind die Handels- und Konsumgüterfirmen, die zu 44 Prozent dabei sind, ML in ersten Projekten und Prototypen zu erproben. Das überrascht insofern nicht, als diese Firmen in der Regel gute gepflegte Datenbestände haben und viel Erfahrung mit Business Intelligence und Data Warehouses besitzen. Gelingt es ihnen, Preisstrategien, Warenverfügbarkeiten oder Marketing-Kampagnen messbar zu verbessern, wird ML als willkommenes Innovationsinstrument bestehender Big-Data-Strategien gesehen.

Gleiches gilt für die IT-, TK- und Medienbranche: Dort kommen ML-Verfahren etwa zum Ausspielen von Online-Werbung, Berechnen von Kaufwahrscheinlichkeiten (Conversion Rates) oder dem Personalisieren von Webinhalten und Einkaufsempfehlungen längst zum Einsatz. Bei den professionellen Dienstleistern spielen das Messen und Verbessern der Kundenbindung, der Dienstleistungsqualität und der Termintreue eine wichtige Rolle, sind das doch die wettbewerbsdifferenzierenden Faktoren.

IT-Abteilungen sind zuständig

Knapp 60 Prozent der befragten Entscheider gaben an, ihre IT-Abteilung sei federführend zuständig, wenn es um ML-Projekte gehe. Den Studienautoren von Crisp zufolge liegt das an der hohen technologischen Komplexität des Themas. Neben mathematischen und statistischen Skills ist demnach auch eine große Bandbreite an Fertigkeiten im Bereich der IT-Operations gefragt. Hinzu kommen die BI- und Analytics-Fähigkeiten, die hier oftmals angesiedelt sind.

Doch auch Fachabteilungen wie Logistik und Produktion sind mit im Boot, weil sie in der Regel die Prozessverbesserungs- und IoT-Szenarien vorantreiben. Die großen Mengen an Maschinen-, Produktions-, Logistik- sowie sonstigen Sensor- und Log-Daten müssen auf Muster und Korrelationen hin abgefragt werden – eine Aufgabe für Fertigung und Logistik.

Und schließlich sind auch Kundenservice und -support führende Instanzen, wenn es um die Einführung von ML geht. Sie wollen die personalisierte Kundeninteraktion vorantreiben und sammeln in ihren Bereichen die Text-, Bild- und Audiodaten, die das Potenzial für Analysen bieten. Interessant an der Umfrage ist indes, dass Marketing und Kommunikation von ML oft nichts wissen wollen, obwohl sie reichlich Einsatzszenarien hätten. Sie könnten etwa Kundenbeziehungen auswerten und die Kundenbindung verbessern, automatisiertes Medien-Monitoring vorantreiben oder das Social Web mit Sentiment-Analysen bearbeiten. All das findet aber relativ selten statt, was Crisp Research mit der traditionell „passiven, technologieagnostischen Rolle“ dieser Abteilungen begründet. Marketing- und Kommunikationsabteilungen treten demnach meist als „Anforderer“ und interne Kunden auf, nicht als diejenigen, die tiefer in Technologien einsteigen.

Welche Machine-Learning-Funktionen benötigen Unternehmen wofür? Und wann kommen welche Lernstile, Frameworks, Programmiersprachen und Algorithmen zum Einsatz? Meistens beginnen Firmen mit Bildanalyse und -erkennung.
  • Bild- und Spracherkennung sind die wichtigsten Anwendungen im Bereich Machine Learning
  • Geht es um die Plattformauswahl, wird die Public Cloud zunehmend wichtig
  • Grafikprozessoren setzen sich im Bereich Deep Learning durch

Wie die Analysten von Crisp Research im Rahmen einer umfassenden Studie gemeinsam mit The unbelievable Machine Company und Hewlett-Packard Enterprise (HPE) schreiben, gibt die Mehrheit der rund 250 befragten IT-Entscheider an, mit der Bildanalyse und -erkennung in das komplexe Thema Machine Learning (ML) einzusteigen. So werden beispielsweise in Industrieunternehmen Fremdkörper auf Förderbändern identifiziert, fehlerhafte Einfärbungen von Produkten entdeckt oder von autonomen Fahrzeugen Straßenschilder erkannt.

Diese Machine-Learning-Funktionen nutzen die Anwender.
Diese Machine-Learning-Funktionen nutzen die Anwender.
Foto: Crisp Research, Kassel


Wichtig sind ML-Verfahren auch zur Sprachsteuerung und -erkennung (42 Prozent). Eng damit verbunden sind Natural Language Processing und Textanalyse – also das semantische Erfassen von Sprachinhalten und Texten. Heute beschäftigen sich 35 Prozent der Unternehmen damit, Tendenz steigend. Hintergrund ist, dass konversationsbasierte Benutzerschnittstellen derzeit einen Aufschwung erleben.

Chatbots, Gesichtserkennung, Sentiment-Analyse und mehr

Machine Learning kommt außerdem bei rund einem Drittel der Befragten im Zusammenhang mit der Entwicklung digitaler Assistenten, sogenannter Bots zum Einsatz. Weitere Einsatzgebiete sind Gesichtserkennung, die Sentiment-Analyse und besondere Verfahren der Mustererkennung – oft in einem unternehmens- oder branchenspezifischen Kontext. Die Spracherkennung ist vor allem für Marketingentscheider interessant, da digitale Assistenten für die Automatisierung von Call-Center-Abläufen oder die Echtzeit-Kommunikation mit dem Kunden an Bedeutung gewinnen. Auch die Personalisierung von Produktempfehlungen ist ein wichtiger Use-Case.

Ein Blick auf die Nutzungsszenarien von ML-Technologien zeigt, dass Bildanalyse und -erkennung heute weit vorne rangieren, doch die Zukunft gehört eher der Sprachsteuerung und – erkennung, ebenso der Textanalyse und Natural Language Processing (NLP). Insgesamt werden ML-Technologien auf breiter Front an Bedeutung gewinne, auch etwa im Bereich der Videoanalyse, der Sentiment-Analyse, der Gesichtserkennung sowie beim Einsatz intelligenter Bots.

Schaut man auf die einzelnen Unternehmensbereiche, so wird deutlich, dass sich die für Customer Experience Management zuständigen Einheiten ML-Technologien vor allem im Bereich der Kundensegmentierung, der personalisierten Produktempfehlung, der Spracherkennung und teilweise auch der Gesichtserkennung bedienen. IT-Abteilungen treiben damit E-Mail-Klassifizierung, Spam-Erkennung, Diagnosesysteme und das Klassifizieren von Dokumenten voran. Die Produktion ist vor allem auf Prozessverbesserungen aus, während Kundendienst und Support ihre Diagnoseysteme vorantreiben und an automatisierten Lösungsempfehlungen arbeiten. Auch Call-Center-Gespräche werden bereits analysiert, teilweise auch mit der Absicht, positive und negative Äußerungen der Kunden zu erkennen (Sentiment-Analyse).

Auch die Bereiche Finance und Human Resources sowie das Management generell nutzen vermehrt ML-Technologien. Wichtigstes Einsatzgebiet sind hier das Risiko-Management sowie Forecasting und Prognosen. Im HR-Bereich werden auch Trainingsempfehlungen automatisiert erstellt, Lebensläufe überprüft und das Talent-Management vorangetrieben. Im zentralen Einkauf und dem Management der Lieferanten ist die Digital Supply-Chain-Verbesserung das Kernaufgabengebiet von ML-technologie. Vermehrt werden hier auch Demand Forecastings ermittelt, Risiken im Zusammenhang mit bestimmten Lieferanten analysiert und generell Entscheidungsprozesse digital unterstützt.

Machine-Learning-Plattformen und -Produkte

Geht es um die Auswahl von Plattformen und -Produkten, spielen Lösungen aus der Public Cloud eine zunehmend wichtige Rolle (Machine Learning as a Service). Um Komplexität aus dem Wege zu gehen und weil die großen Cloud-Provider auch die maßgeblichen Innovatoren auf diesem Gebiet sind, entscheiden sich viele Anwender für diese Cloud-Lösungen. Während 38,1 der Befragten Lösungen aus der Public-Cloud bevorzugen, wählen 19,1 Prozent proprietäre Lösungen ausgesuchter Anbieter und 18,5 Prozent Open-Source-Alternativen. Der Rest verfolgt entweder eine hybride Strategie (15,5 Prozent) oder hat sich noch keine Meinung dazu gebildet (8,8 Prozent).

Welche Cloud-Angebote zu Machine Learning sind im Einsatz?
Welche Cloud-Angebote zu Machine Learning sind im Einsatz?
Foto: Crisp Research


Unter den Cloud-basierten Lösungen hat AWS den höchsten Bekanntheitsgrad: 71 Prozent der Entscheider geben an, dass ihnen Amazon in diesem Kontext bekannt sei. Auch Microsoft, Google und IBM sind den Umfrageteilnehmern zu mehr als zwei Drittel im ML-Umfeld ein Begriff. Interessanterweise nutzen aber nur 17 Prozent der befragten die AWS-Cloud-Dienste im Kontext der Evaluierung, Projektierung sowie im produktiven Betrieb für ML. Jeweils rund ein Drittel der Befragten beschäftigt sich indes mit IBM Watson, Microsoft Azure oder der Google Cloud Machine Learning Plattform.

Die Analysten nehmen an, dass dies viel mit den Marketing-Anstrengungen der Hersteller zu tun hat. IBM und Microsoft investieren demnach massiv in ihre Cognitive- beziehungsweise KI-Strategie. Beide haben einen starken Mittelstands- und Großkundenvertrieb und ein großes Partnernetzwerk. Google indes verdanke seine Position dem Image als gewaltige daten- und Analytics-Maschine, die den Markt durch viele Innovationen treibe – etwa Tensorflow, viele ML-APIs und auch eigene Hardware. Schließlich zähle aber auch HP Enterprise mit „Haven on Demand“ zu den relevanten ML-Playern und werde von 14 Prozent der Befragten genutzt.

Deep Learning ist schwieriger

Bereits in den 40er Jahren des vergangenen Jahrhunderts wurden die ersten neuronalen Lernregeln beschrieben. Die wissenschaftlichen Erkenntnisse wuchsen rasch, die Anzahl der Algorithmen ebenfalls – doch es fehlte an der notwendigen Rechenleistung, um „Rückgekoppelte Neuronale Netzwerke“ in der Fläche zu nutzen. Heute sind diese unter dem Begriff Deep Learning in aller Munde, sie könnten Bereiche wie Handschriftenerkennung, Spracherkennung, maschinelles Übersetzen oder auch automatische Bildbeschreibungen revolutionieren.

Hintergrund ist, dass eine Präzision erreicht werden kann, die menschliche Fähigkeiten im jeweiligen Zusammenhang weit übertrifft. Dabei spannen neuronale Netze Ebenen von unterschiedlicher Komplexität auf. Je mehr Daten so einem neuronalen Netz zum Trainieren zur Verfügung stehen, desto besser werden die Ergebnisse beziehungsweise die trainierte Künstliche Intelligenz. So lernt ein System beispielsweise, wie anhand einer Computer-Tomografie Krebsgeschwüre diagnostiziert werden können, die das menschliche Auge nicht so einfach sieht.

Grafikprozessoren bieten die nötige Performance

Im Bereich des Deep Learning haben sich hardwareseitig Grafikprozessoren (GPUs) wegen ihre hohen Performance als besonders geeignet erwiesen. Förderlich waren außerdem die schier unbegrenzte Rechenpower, die sich aus den Public-Cloud-Ressourcen ergibt, sowie die Verfügbarkeit großer Mengen von Daten aus den verschiedensten Anwendungsgebieten. Unternehmen nutzen bereits Deep-learning-Algorithmen, im bestimmte Merkmal in Bildern aufzuspüren, Videoanalysen vorzunehmen, Umweltparameter beim autonomen Fahren zu verarbeiten oder automatische Sprachverarbeitung voranzutreiben.

In der Crisp-Umfrage geben 48 Prozent der Teilnehmer an, von Deep Learning zumindest gehört oder gelesen zu haben. Weitere 21 Prozent sind bereits in einer konkreten Evaluationsphase. Sie haben Erkenntnisse gesammelt und arbeiten nun an konkreten Prototypen, um ihr gewünschtes Einsatzszenario zu validieren. Weitere fünf Prozent sind sogar noch einen Schritt weiter und haben bereits Deep Learning im Einsatz. Vor allem Startups und Konzerne – auch hier wieder vor allem aus dem Automotive-Sektor – haben hier die Nase vorn.

Unter den Frameworks und Bibliotheken, die für das Implementieren von Deep-Learning-Algorithmen eine Rolle spielen, spielen unter anderem Microsofts „Computational Network Toolkit“ (CNTK) sowie jede Menge Public-Cloud- und Open-Source-Lösungen eine Rolle (eine Übersicht gibt es hier

Machine Learning macht Analysen besser

Zuerst analysierten lernende Maschinen das Nutzerverhalten in Suchmaschinen, um passende Werbung anzuzeigen. Heute optimieren sie Verkehrsflüsse, die Stahlherstellung und planen die Flugzeugwartung. Experten von Allianz, Trip Advisor, GfK und Boeing erklären, wie ihnen Machine Learning hilft.,3217540

Bei der Münchener Allianz Versicherung ist Andreas Braun, Head of Global Data and Analytics, zufrieden mit den Ergebnissen seiner Experimente mit den neuen Analytics-Ansätzen aus der künstlichen Intelligenz. „Wir haben bei uns ein Ökosystem aus verschiedenen Bestandteilen im Einsatz. Big-Data-Technologien und Machine Learning bieten uns bessere Möglichkeiten, mit unseren Daten umzugehen, und liefern konsistent gute Ergebnisse“, sagte er auf der Konferenz der Yandex Data Factory zum Thema „Machine Learning and Big Data“ in Berlin. Zum Beispiel im Gebäude-Management: Zusammen mit Studenten der TU München hat die Versicherung eine App entwickelt, die eine Vielzahl von Gegenständen über Sensoren vernetzt.

„Das System kalibriert sich selbst, lernt normales Verhalten im Haus, und kann so einen Einbruch von anderen ungewöhnlichen, aber unkritischen Vorfällen unterscheiden.“ Außerdem wollen die Experten die Bilderkennung weiter verbessern. Eingereichte Fotos sollen bei Versicherungsschäden automatisch durch Maschinen beurteilt werden.

Die Experten, die der russische Suchmaschinen-Anbieter Yandex nach Berlin eingeladen hatte, tauschten sich unter dem Motto „Business Challenges“ auch über die Schwierigkeiten und Risiken rund um Machine Learning aus. Jeff Palmucci, Director of Machine Intelligence beim Reiseportal Trip Advisor, schilderte, wie sein Unternehmen maschinelles Lernen in die Geschäftsprozesse implementiert. So hilft die Technik, Restaurants und Hotels automatisiert mit passenden Tags wie „romantisch“ oder „charmant“ zu versehen, damit Suchende schnell das richtige Angebot finden. Auch um Betrug etwa bei den Bewertungen rasch zu erkennen, setzt das Portal Machine Learning ein.

Menschliches Verhalten vorhersagen

Machine Learning stellt Unternehmen vor vielfältige Herausforderungen. Nicht alle Branchen eignen sich gleich gut, erklärte Jane Zavalishina, CEO der Yandex Data Factory: „Es geht vor allem darum, menschliches Verhalten vorherzusagen.“ Bei Ergebnissen, die auf Machine Learning basieren, könne man aber durch die hohe Komplexität und die großen Datenmengen nie genau nachvollziehen, wie sie zustande gekommen sind. In der Praxis müsse man mit den Empfehlungen experimentieren, um herauszufinden, ob sie der bisherigen Vorgehensweise überlegen sind. Das gehe aus ethischen und praktischen Gründen allerdings nicht immer.

Jane Zavalishina CEO, Yandex Data Factory „Viele Unternehmen befinden sich aber noch an dem Punkt, an dem sie versuchen, Big Data Analytics überhaupt zu verstehen.“
Jane Zavalishina CEO, Yandex Data Factory „Viele Unternehmen befinden sich aber noch an dem Punkt, an dem sie versuchen, Big Data Analytics überhaupt zu verstehen.“
Foto: Yandex

In Echtzeit Web-Inhalte zu personalisieren oder Vorhersagen zu treffen, ist für die russische Suchmaschine Yandex nichts Neues. Das Wissen des Konzerns, das aus der Suchtechnik und dem kontextuellen Einspielen passender Werbung entstanden ist, und die dafür entwickelten Algorithmen stellt sie seit 2014 auch extern zur Verfügung. Zunächst probierte das Tochterunternehmen Yandex Data Factory, das Firmensitze in Moskau und Amsterdam unterhält, die Techniken maschinellen Lernens in der Wissenschaft aus – zum Beispiel, um Big-Data-Probleme des europäischen Kernforschungszentrums CERN zu lösen.

Inzwischen besprechen die Datenexperten mit Firmen, die viele Kunden und große Datenmengen haben, wie sich deren Services, Prozesse und Produkte ver­bessern lassen. „Die Anwendungsmöglichkeiten für maschinelles Lernen in Unternehmen sind fast unbegrenzt“, sagte Zavalishina. „Viele Unternehmen befin­den sich aber noch an dem Punkt, an dem sie versuchen, Big Data Analytics überhaupt zu verstehen.“

Eine der ersten Firmen, die Wissen und Technologie von Yandex nutzte, war die russische Straßenverwaltungsbehörde Rosavtodor, die Vorhersagen zur Verkehrsdichte und zu Unfällen benötigte. Im Stahlwerk Magnitogorsk Iron and Steel Works optimieren heute Algorithmen die Stahlproduktion. Zu wenige Zusätze ergeben eine schlechte Qualität, zu viele treiben die Kosten in die Höhe. Bisher nutzten die Stahlkocher für ihre Mischungsvorhersagen komplizierte Modelle. Yandex Data Factory verwendete zur Optimierung historische Daten aus den zurückliegenden zehn Jahren. Vergleichsweise einfach scheint es dagegen, mit Machine Learning Websites zu optimieren und Online-Werbung auszusenden.

Business ist datengetrieben

„Wir sind ein komplett datengetriebenes Business“, sagt Norbert Wirth, Global Head of Data and Science beim Marktforschungsinstitut GfK, „Machine-Learning-Algorithmen sind für uns ein Werkzeug im Kanon mit anderen, das aber für die Vorhersage und für Klassifizierungsprobleme zunehmend wichtiger wird.“ GfK nutzt es derzeit vor allem für die Analyse von Social-Media-Daten und um Marktanteile und Marktperformance vorherzusagen.

„Wir setzen es ein, wenn nicht die Frage nach dem Warum entscheidend ist, sondern die Qualität der Vorhersage“, so Wirth. Sind Aussagen über eine Marke tendenziell eher positiv oder negativ? Und um welche Themen geht es? Bei kleineren Datenbeständen könne man das noch selbst herausfinden, wird es jedoch umfangreicher, seien die Algorithmen „extrem spannend – und sie werden immer leistungsfähiger“. Das sei kein Hype, sagt der Marktforscher, „Machine Learning wird an Bedeutung zunehmen. Mit wachsender Computerpower kann man damit jetzt wirklich arbeiten.“ Die eine Sache sei ein toller Algorithmus, die andere, ob man die dafür nötigen Maschinen auch am Start habe.

In Zukunft werden Analysten laut Wirth zusätzliche Daten verwenden, um Algorithmen zu trainieren und die Modelle leistungsfähiger zu machen. „Es geht in die Richtung, im Analyseprozess mit mehreren Datenquellen zu arbeiten. Natürlich mit solchen, die auch legal genutzt werden dürfen.“ Data Privacy sei ein sehr wichtiges Thema rund um Machine Learning – aber auch die Stabilität und die Qualität der Daten.

Der Flugzeughersteller Boeing nutzt Machine Learning, um seine Services und die interne Produktion zu verbessern, berichtete Sergey Kravchenko, President Russia and CIS von Boeing. Das Flugzeug 787 verfüge über mehr als zehntausend mit dem Internet verbundene Sensoren, die den Mechanikern am Boden schon während des Fluges melden, wenn zum Beispiel eine Lampe oder eine Pumpe ausgetauscht werden muss. So können Fluggesellschaften ihre Wartungskosten reduzieren und im Betrieb effizienter arbeiten.

Boeing arbeitet mit Big Data und Machine Learning, um den Fluggesellschaften mit den während eines Flugs gesammelten Daten zu helfen, Treibstoffkosten zu senken und die Piloten bei schlechtem Wetter zu unterstützen. Nun werden die Daten auch in der Produktion verwendet, um etwa für bestimmte Prozesse die besten Ingenieure zu finden. Daten der Personalabteilung würden genutzt, um zu verstehen, wie die Lebensdauer und die Qualität der Flugzeuge mit dem Training und der Mischung der Menschen im Produktionsteam korrelieren. Gibt es bei Prozessen, die aufwendige Nacharbeiten erfordern, Zusammenhänge mit den bereitgestellten Werkzeugen oder mit dem Team? Kravchenko will mit Big-Data-Analysen den gesamten Zyklus von Design, Produktion und Wartung verbessern.

Ein neues Big-Data-Projekt ist die Flight Training Academy, die 2016 eröffnet werden soll. Hier werden Daten der drei Flugsimulatoren gesammelt und ausgewertet, um die Gestaltung des Cockpits und das Design der Flugzeug­software zu verbessern. Kravchenko will seinen russischen Kunden auch anbieten, in Zukunft Daten auszutauschen und sie gemeinsam auszuwerten.

Experten müssen zusammenpassen

Die Fertigungsindustrie stehe bei der Anwendung von Machine Learning – verglichen etwa mit Telcos und dem Handel – noch am Anfang. Sie werde aber schnell von ihnen und auch von Firmen wie Amazon und Google, lernen. Wer Erfolg haben wolle, müsse die besten Flugzeug- und IT-Experten zusammenbringen. Das Problem: „Die kommen von verschiedenen Planeten.“

Die Zusammenarbeit kann dennoch gelingen – wenn sich alle auf eine gemeinsame Terminologie einigen. „Die Datenexperten müssen etwas mehr von Flugzeugen und Airlines verstehen und die Flugzeugspezialisten mehr über Data Analytics lernen. Sie müssen sich die Werkzeuge teilen, sich gegenseitig vertrauen und ein gemeinsames Team aufbauen“, sagt der Flugzeugbauer. Ein weiteres Problem sei die Relevanz der Daten. „Hier muss die Industrie ihre riesigen Datenmengen anschauen und entscheiden, welche Daten wirklich wichtig sind, um bestimmte Probleme zu lösen. Das ist nicht einfach, dafür brauchen wir Zeit, Trial and Error, und wir müssen von anderen Branchen lernen.“ Die richtige Auswahl der Daten und die Interpretation der Ergebnisse seien dabei wichtiger als der Algorithmus selbst.