Tag Archives: experiment
#433758 DeepMind’s New Research Plan to Make ...
Making sure artificial intelligence does what we want and behaves in predictable ways will be crucial as the technology becomes increasingly ubiquitous. It’s an area frequently neglected in the race to develop products, but DeepMind has now outlined its research agenda to tackle the problem.
AI safety, as the field is known, has been gaining prominence in recent years. That’s probably at least partly down to the overzealous warnings of a coming AI apocalypse from well-meaning, but underqualified pundits like Elon Musk and Stephen Hawking. But it’s also recognition of the fact that AI technology is quickly pervading all aspects of our lives, making decisions on everything from what movies we watch to whether we get a mortgage.
That’s why DeepMind hired a bevy of researchers who specialize in foreseeing the unforeseen consequences of the way we built AI back in 2016. And now the team has spelled out the three key domains they think require research if we’re going to build autonomous machines that do what we want.
In a new blog designed to provide updates on the team’s work, they introduce the ideas of specification, robustness, and assurance, which they say will act as the cornerstones of their future research. Specification involves making sure AI systems do what their operator intends; robustness means a system can cope with changes to its environment and attempts to throw it off course; and assurance involves our ability to understand what systems are doing and how to control them.
A classic thought experiment designed to illustrate how we could lose control of an AI system can help illustrate the problem of specification. Philosopher Nick Bostrom’s posited a hypothetical machine charged with making as many paperclips as possible. Because the creators fail to add what they might assume are obvious additional goals like not harming people, the AI wipes out humanity so we can’t switch it off before turning all matter in the universe into paperclips.
Obviously the example is extreme, but it shows how a poorly-specified goal can lead to unexpected and disastrous outcomes. Properly codifying the desires of the designer is no easy feat, though; often there are not neat ways to encompass both the explicit and implicit goals in ways that are understandable to the machine and don’t leave room for ambiguities, meaning we often rely on incomplete approximations.
The researchers note recent research by OpenAI in which an AI was trained to play a boat-racing game called CoastRunners. The game rewards players for hitting targets laid out along the race route. The AI worked out that it could get a higher score by repeatedly knocking over regenerating targets rather than actually completing the course. The blog post includes a link to a spreadsheet detailing scores of such examples.
Another key concern for AI designers is making their creation robust to the unpredictability of the real world. Despite their superhuman abilities on certain tasks, most cutting-edge AI systems are remarkably brittle. They tend to be trained on highly-curated datasets and so can fail when faced with unfamiliar input. This can happen by accident or by design—researchers have come up with numerous ways to trick image recognition algorithms into misclassifying things, including thinking a 3D printed tortoise was actually a gun.
Building systems that can deal with every possible encounter may not be feasible, so a big part of making AIs more robust may be getting them to avoid risks and ensuring they can recover from errors, or that they have failsafes to ensure errors don’t lead to catastrophic failure.
And finally, we need to have ways to make sure we can tell whether an AI is performing the way we expect it to. A key part of assurance is being able to effectively monitor systems and interpret what they’re doing—if we’re basing medical treatments or sentencing decisions on the output of an AI, we’d like to see the reasoning. That’s a major outstanding problem for popular deep learning approaches, which are largely indecipherable black boxes.
The other half of assurance is the ability to intervene if a machine isn’t behaving the way we’d like. But designing a reliable off switch is tough, because most learning systems have a strong incentive to prevent anyone from interfering with their goals.
The authors don’t pretend to have all the answers, but they hope the framework they’ve come up with can help guide others working on AI safety. While it may be some time before AI is truly in a position to do us harm, hopefully early efforts like these will mean it’s built on a solid foundation that ensures it is aligned with our goals.
Image Credit: cono0430 / Shutterstock.com Continue reading
Posted in Human Robots
#432467 Dungeons and Dragons, Not Chess and Go: ...
Everyone had died—not that you’d know it, from how they were laughing about their poor choices and bad rolls of the dice. As a social anthropologist, I study how people understand artificial intelligence (AI) and our efforts towards attaining it; I’m also a life-long fan of Dungeons and Dragons (D&D), the inventive fantasy roleplaying game. During a recent quest, when I was playing an elf ranger, the trainee paladin (or holy knight) acted according to his noble character, and announced our presence at the mouth of a dragon’s lair. The results were disastrous. But while success in D&D means “beating the bad guy,” the game is also a creative sandbox, where failure can count as collective triumph so long as you tell a great tale.
What does this have to do with AI? In computer science, games are frequently used as a benchmark for an algorithm’s “intelligence.” The late Robert Wilensky, a professor at the University of California, Berkeley and a leading figure in AI, offered one reason why this might be. Computer scientists “looked around at who the smartest people were, and they were themselves, of course,” he told the authors of Compulsive Technology: Computers as Culture (1985). “They were all essentially mathematicians by training, and mathematicians do two things—they prove theorems and play chess. And they said, hey, if it proves a theorem or plays chess, it must be smart.” No surprise that demonstrations of AI’s “smarts” have focused on the artificial player’s prowess.
Yet the games that get chosen—like Go, the main battlefield for Google DeepMind’s algorithms in recent years—tend to be tightly bounded, with set objectives and clear paths to victory or defeat. These experiences have none of the open-ended collaboration of D&D. Which got me thinking: do we need a new test for intelligence, where the goal is not simply about success, but storytelling? What would it mean for an AI to “pass” as human in a game of D&D? Instead of the Turing test, perhaps we need an elf ranger test?
Of course, this is just a playful thought experiment, but it does highlight the flaws in certain models of intelligence. First, it reveals how intelligence has to work across a variety of environments. D&D participants can inhabit many characters in many games, and the individual player can “switch” between roles (the fighter, the thief, the healer). Meanwhile, AI researchers know that it’s super difficult to get a well-trained algorithm to apply its insights in even slightly different domains—something that we humans manage surprisingly well.
Second, D&D reminds us that intelligence is embodied. In computer games, the bodily aspect of the experience might range from pressing buttons on a controller in order to move an icon or avatar (a ping-pong paddle; a spaceship; an anthropomorphic, eternally hungry, yellow sphere), to more recent and immersive experiences involving virtual-reality goggles and haptic gloves. Even without these add-ons, games can still produce biological responses associated with stress and fear (if you’ve ever played Alien: Isolation you’ll understand). In the original D&D, the players encounter the game while sitting around a table together, feeling the story and its impact. Recent research in cognitive science suggests that bodily interactions are crucial to how we grasp more abstract mental concepts. But we give minimal attention to the embodiment of artificial agents, and how that might affect the way they learn and process information.
Finally, intelligence is social. AI algorithms typically learn through multiple rounds of competition, in which successful strategies get reinforced with rewards. True, it appears that humans also evolved to learn through repetition, reward and reinforcement. But there’s an important collaborative dimension to human intelligence. In the 1930s, the psychologist Lev Vygotsky identified the interaction of an expert and a novice as an example of what became called “scaffolded” learning, where the teacher demonstrates and then supports the learner in acquiring a new skill. In unbounded games, this cooperation is channelled through narrative. Games of It among small children can evolve from win/lose into attacks by terrible monsters, before shifting again to more complex narratives that explain why the monsters are attacking, who is the hero, and what they can do and why—narratives that aren’t always logical or even internally compatible. An AI that could engage in social storytelling is doubtless on a surer, more multifunctional footing than one that plays chess; and there’s no guarantee that chess is even a step on the road to attaining intelligence of this sort.
In some ways, this failure to look at roleplaying as a technical hurdle for intelligence is strange. D&D was a key cultural touchstone for technologists in the 1980s and the inspiration for many early text-based computer games, as Katie Hafner and Matthew Lyon point out in Where Wizards Stay up Late: The Origins of the Internet (1996). Even today, AI researchers who play games in their free time often mention D&D specifically. So instead of beating adversaries in games, we might learn more about intelligence if we tried to teach artificial agents to play together as we do: as paladins and elf rangers.
This article was originally published at Aeon and has been republished under Creative Commons.
Image Credit:Benny Mazur/Flickr / CC BY 2.0 Continue reading
Posted in Human Robots
#432051 What Roboticists Are Learning From Early ...
You might not have heard of Hanson Robotics, but if you’re reading this, you’ve probably seen their work. They were the company behind Sophia, the lifelike humanoid avatar that’s made dozens of high-profile media appearances. Before that, they were the company behind that strange-looking robot that seemed a bit like Asimo with Albert Einstein’s head—or maybe you saw BINA48, who was interviewed for the New York Times in 2010 and featured in Jon Ronson’s books. For the sci-fi aficionados amongst you, they even made a replica of legendary author Philip K. Dick, best remembered for having books with titles like Do Androids Dream of Electric Sheep? turned into films with titles like Blade Runner.
Hanson Robotics, in other words, with their proprietary brand of life-like humanoid robots, have been playing the same game for a while. Sometimes it can be a frustrating game to watch. Anyone who gives the robot the slightest bit of thought will realize that this is essentially a chat-bot, with all the limitations this implies. Indeed, even in that New York Times interview with BINA48, author Amy Harmon describes it as a frustrating experience—with “rare (but invariably thrilling) moments of coherence.” This sensation will be familiar to anyone who’s conversed with a chatbot that has a few clever responses.
The glossy surface belies the lack of real intelligence underneath; it seems, at first glance, like a much more advanced machine than it is. Peeling back that surface layer—at least for a Hanson robot—means you’re peeling back Frubber. This proprietary substance—short for “Flesh Rubber,” which is slightly nightmarish—is surprisingly complicated. Up to thirty motors are required just to control the face; they manipulate liquid cells in order to make the skin soft, malleable, and capable of a range of different emotional expressions.
A quick combinatorial glance at the 30+ motors suggests that there are millions of possible combinations; researchers identify 62 that they consider “human-like” in Sophia, although not everyone agrees with this assessment. Arguably, the technical expertise that went into reconstructing the range of human facial expressions far exceeds the more simplistic chat engine the robots use, although it’s the second one that allows it to inflate the punters’ expectations with a few pre-programmed questions in an interview.
Hanson Robotics’ belief is that, ultimately, a lot of how humans will eventually relate to robots is going to depend on their faces and voices, as well as on what they’re saying. “The perception of identity is so intimately bound up with the perception of the human form,” says David Hanson, company founder.
Yet anyone attempting to design a robot that won’t terrify people has to contend with the uncanny valley—that strange blend of concern and revulsion people react with when things appear to be creepily human. Between cartoonish humanoids and genuine humans lies what has often been a no-go zone in robotic aesthetics.
The uncanny valley concept originated with roboticist Masahiro Mori, who argued that roboticists should avoid trying to replicate humans exactly. Since anything that wasn’t perfect, but merely very good, would elicit an eerie feeling in humans, shirking the challenge entirely was the only way to avoid the uncanny valley. It’s probably a task made more difficult by endless streams of articles about AI taking over the world that inexplicably conflate AI with killer humanoid Terminators—which aren’t particularly likely to exist (although maybe it’s best not to push robots around too much).
The idea behind this realm of psychological horror is fairly simple, cognitively speaking.
We know how to categorize things that are unambiguously human or non-human. This is true even if they’re designed to interact with us. Consider the popularity of Aibo, Jibo, or even some robots that don’t try to resemble humans. Something that resembles a human, but isn’t quite right, is bound to evoke a fear response in the same way slightly distorted music or slightly rearranged furniture in your home will. The creature simply doesn’t fit.
You may well reject the idea of the uncanny valley entirely. David Hanson, naturally, is not a fan. In the paper Upending the Uncanny Valley, he argues that great art forms have often resembled humans, but the ultimate goal for humanoid roboticists is probably to create robots we can relate to as something closer to humans than works of art.
Meanwhile, Hanson and other scientists produce competing experiments to either demonstrate that the uncanny valley is overhyped, or to confirm it exists and probe its edges.
The classic experiment involves gradually morphing a cartoon face into a human face, via some robotic-seeming intermediaries—yet it’s in movement that the real horror of the almost-human often lies. Hanson has argued that incorporating cartoonish features may help—and, sometimes, that the uncanny valley is a generational thing which will melt away when new generations grow used to the quirks of robots. Although Hanson might dispute the severity of this effect, it’s clearly what he’s trying to avoid with each new iteration.
Hiroshi Ishiguro is the latest of the roboticists to have dived headlong into the valley.
Building on the work of pioneers like Hanson, those who study human-robot interaction are pushing at the boundaries of robotics—but also of social science. It’s usually difficult to simulate what you don’t understand, and there’s still an awful lot we don’t understand about how we interpret the constant streams of non-verbal information that flow when you interact with people in the flesh.
Ishiguro took this imitation of human forms to extreme levels. Not only did he monitor and log the physical movements people made on videotapes, but some of his robots are based on replicas of people; the Repliee series began with a ‘replicant’ of his daughter. This involved making a rubber replica—a silicone cast—of her entire body. Future experiments were focused on creating Geminoid, a replica of Ishiguro himself.
As Ishiguro aged, he realized that it would be more effective to resemble his replica through cosmetic surgery rather than by continually creating new casts of his face, each with more lines than the last. “I decided not to get old anymore,” Ishiguro said.
We love to throw around abstract concepts and ideas: humans being replaced by machines, cared for by machines, getting intimate with machines, or even merging themselves with machines. You can take an idea like that, hold it in your hand, and examine it—dispassionately, if not without interest. But there’s a gulf between thinking about it and living in a world where human-robot interaction is not a field of academic research, but a day-to-day reality.
As the scientists studying human-robot interaction develop their robots, their replicas, and their experiments, they are making some of the first forays into that world. We might all be living there someday. Understanding ourselves—decrypting the origins of empathy and love—may be the greatest challenge to face. That is, if you want to avoid the valley.
Image Credit: Anton Gvozdikov / Shutterstock.com Continue reading
Posted in Human Robots