More stories

  • in

    After cracking the 'sum of cubes' puzzle for 42, mathematicians discover a new solution for 3

    What do you do after solving the answer to life, the universe, and everything? If you’re mathematicians Drew Sutherland and Andy Booker, you go for the harder problem.
    In 2019, Booker, at the University of Bristol, and Sutherland, principal research scientist at MIT, were the first to find the answer to 42. The number has pop culture significance as the fictional answer to “the ultimate question of life, the universe, and everything,” as Douglas Adams famously penned in his novel “The Hitchhiker’s Guide to the Galaxy.” The question that begets 42, at least in the novel, is frustratingly, hilariously unknown.
    In mathematics, entirely by coincidence, there exists a polynomial equation for which the answer, 42, had similarly eluded mathematicians for decades. The equation x3+y3+z3=k is known as the sum of cubes problem. While seemingly straightforward, the equation becomes exponentially difficult to solve when framed as a “Diophantine equation” — a problem that stipulates that, for any value of k, the values for x, y, and z must each be whole numbers.
    When the sum of cubes equation is framed in this way, for certain values of k, the integer solutions for x, y, and z can grow to enormous numbers. The number space that mathematicians must search across for these numbers is larger still, requiring intricate and massive computations.
    Over the years, mathematicians had managed through various means to solve the equation, either finding a solution or determining that a solution must not exist, for every value of k between 1 and 100 — except for 42.
    In September 2019, Booker and Sutherland, harnessing the combined power of half a million home computers around the world, for the first time found a solution to 42. The widely reported breakthrough spurred the team to tackle an even harder, and in some ways more universal problem: finding the next solution for 3.

    advertisement

    Booker and Sutherland have now published the solutions for 42 and 3, along with several other numbers greater than 100, this week in the Proceedings of the National Academy of Sciences.
    Picking up the gauntlet
    The first two solutions for the equation x3+y3+z3 = 3 might be obvious to any high school algebra student, where x, y, and z can be either 1, 1, and 1, or 4, 4, and -5. Finding a third solution, however, has stumped expert number theorists for decades, and in 1953 the puzzle prompted pioneering mathematician Louis Mordell to ask the question: Is it even possible to know whether other solutions for 3 exist?
    “This was sort of like Mordell throwing down the gauntlet,” says Sutherland. “The interest in solving this question is not so much for the particular solution, but to better understand how hard these equations are to solve. It’s a benchmark against which we can measure ourselves.”
    As decades went by with no new solutions for 3, many began to believe there were none to be found. But soon after finding the answer to 42, Booker and Sutherland’s method, in a surprisingly short time, turned up the next solution for 3:5699368212219623807203 + (−569936821113563493509)3 + (−472715493453327032)3 = 3
    The discovery was a direct answer to Mordell’s question: Yes, it is possible to find the next solution to 3, and what’s more, here is that solution. And perhaps more universally, the solution, involving gigantic, 21-digit numbers that were not possible to sift out until now, suggests that there are more solutions out there, for 3, and other values of k.

    advertisement

    “There had been some serious doubt in the mathematical and computational communities, because [Mordell’s question] is very hard to test,” Sutherland says. “The numbers get so big so fast. You’re never going to find more than the first few solutions. But what I can say is, having found this one solution, I’m convinced there are infinitely many more out there.”
    A solution’s twist
    To find the solutions for both 42 and 3, the team started with an existing algorithm, or a twisting of the sum of cubes equation into a form they believed would be more manageable to solve:
    k − z3 = x3 + y3 = (x + y)(x2 − xy + y2)
    This approach was first proposed by mathematician Roger Heath-Brown, who conjectured that there should be infinitely many solutions for every suitable k. The team further modified the algorithm by representing x+y as a single parameter, d. They then reduced the equation by dividing both sides by d and keeping only the remainder — an operation in mathematics termed “modulo d” — leaving a simplified representation of the problem.
    “You can now think of k as a cube root of z, modulo d,” Sutherland explains. “So imagine working in a system of arithmetic where you only care about the remainder modulo d, and we’re trying to compute a cube root of k.”
    With this sleeker version of the equation, the researchers would only need to look for values of d and z that would guarantee finding the ultimate solutions to x, y, and z, for k=3. But still, the space of numbers that they would have to search through would be infinitely large.
    So, the researchers optimized the algorithm by using mathematical “sieving” techniques to dramatically cut down the space of possible solutions for d.
    “This involves some fairly advanced number theory, using the structure of what we know about number fields to avoid looking in places we don’t need to look,” Sutherland says.
    A global task
    The team also developed ways to efficiently split the algorithm’s search into hundreds of thousands of parallel processing streams. If the algorithm were run on just one computer, it would have taken hundreds of years to find a solution to k=3. By dividing the job into millions of smaller tasks, each independently run on a separate computer, the team could further speed up their search.
    In September 2019, the researchers put their plan in play through Charity Engine, a project that can be downloaded as a free app by any personal computer, and which is designed to harness any spare home computing power to collectively solve hard mathematical problems. At the time, Charity Engine’s grid comprised over 400,000 computers around the world, and Booker and Sutherland were able to run their algorithm on the network as a test of Charity Engine’s new software platform.
    “For each computer in the network, they are told, ‘your job is to look for d’s whose prime factor falls within this range, subject to some other conditions,'” Sutherland says. “And we had to figure out how to divide the job up into roughly 4 million tasks that would each take about three hours for a computer to complete.”
    Very quickly, the global grid returned the very first solution to k=42, and just two weeks later, the researchers confirmed they had found the third solution for k=3 — a milestone that they marked, in part, by printing the equation on t-shirts.
    The fact that a third solution to k=3 exists suggests that Heath-Brown’s original conjecture was right and that there are infinitely more solutions beyond this newest one. Heath-Brown also predicts the space between solutions will grow exponentially, along with their searches. For instance, rather than the third solution’s 21-digit values, the fourth solution for x, y, and z will likely involve numbers with a mind-boggling 28 digits.
    “The amount of work you have to do for each new solution grows by a factor of more than 10 million, so the next solution for 3 will need 10 million times 400,000 computers to find, and there’s no guarantee that’s even enough,” Sutherland says. “I don’t know if we’ll ever know the fourth solution. But I do believe it’s out there.” More

  • in

    How to make all headphones intelligent

    How do you turn “dumb” headphones into smart ones? Rutgers engineers have invented a cheap and easy way by transforming headphones into sensors that can be plugged into smartphones, identify their users, monitor their heart rates and perform other services.
    Their invention, called HeadFi, is based on a small plug-in headphone adapter that turns a regular headphone into a sensing device. Unlike smart headphones, regular headphones lack sensors. HeadFi would allow users to avoid having to buy a new pair of smart headphones with embedded sensors to enjoy sensing features.
    “HeadFi could turn hundreds of millions of existing, regular headphones worldwide into intelligent ones with a simple upgrade,” said Xiaoran Fan, a HeadFi primary inventor. He is a recent Rutgers doctoral graduate who completed the research during his final year at the university and now works at Samsung Artificial Intelligence Center.
    A peer-reviewed Rutgers-led paper on the invention, which results in “earable intelligence,” will be formally published in October at MobiCom 2021, the top international conference on mobile computing and mobile and wireless networking.
    Headphones are among the most popular wearable devices worldwide and they continue to become more intelligent as new functions appear, such as touch-based gesture control, the paper notes. Such functions usually rely on auxiliary sensors, such as accelerometers, gyroscopes and microphones that are available on many smart headphones.
    HeadFi turns the two drivers already inside all headphones into a versatile sensor, and it works by connecting headphones to a pairing device, such as a smartphone. It does not require adding auxiliary sensors and avoids changes to headphone hardware or the need to customize headphones, both of which may increase their weight and bulk. By plugging into HeadFi, a converted headphone can perform sensing tasks and play music at the same time.
    The engineers conducted experiments with 53 volunteers using 54 pairs of headphones with estimated prices ranging from $2.99 to $15,000. HeadFi can achieve 97.2 percent to 99.5 percent accuracy on user identification, 96.8 percent to 99.2 percent on heart rate monitoring and 97.7 percent to 99.3 percent on gesture recognition.

    Story Source:
    Materials provided by Rutgers University. Note: Content may be edited for style and length. More

  • in

    Read to succeed — in math; study shows how reading skill shapes more than just reading

    A University at Buffalo researcher’s recent work on dyslexia has unexpectedly produced a startling discovery which clearly demonstrates how the cooperative areas of the brain responsible for reading skill are also at work during apparently unrelated activities, such as multiplication.
    Though the division between literacy and math is commonly reflected in the division between the arts and sciences, the findings suggest that reading, writing and arithmetic, the foundational skills informally identified as the three Rs, might actually overlap in ways not previously imagined, let alone experimentally validated.
    “These findings floored me,” said Christopher McNorgan, PhD, the paper’s author and an assistant professor in UB’s Department of Psychology. “They elevate the value and importance of literacy by showing how reading proficiency reaches across domains, guiding how we approach other tasks and solve other problems.
    “Reading is everything, and saying so is more than an inspirational slogan. It’s now a definitive research conclusion.”
    And it’s a conclusion that was not originally part of McNorgan’s design. He planned to exclusively explore if it was possible to identify children with dyslexia on the basis of how the brain was wired for reading.
    “It seemed plausible given the work I had recently finished, which identified a biomarker for ADHD,” said McNorgan, an expert in neuroimaging and computational modeling.

    advertisement

    Like that previous study, a novel deep learning approach that makes multiple simultaneous classifications is at the core of McNorgan’s current paper, which appears in the journal Frontiers in Computational Neuroscience.
    Deep learning networks are ideal for uncovering conditional, non-linear relationships.
    Where linear relationships involve one variable directly influencing another, a non-linear relationship can be slippery because changes in one area do not necessarily proportionally influence another area. But what’s challenging for traditional methods is easily handled through deep learning.
    McNorgan identified dyslexia with 94% accuracy when he finished with his first data set, consisting of functional connectivity from 14 good readers and 14 poor readers engaged in a language task.
    But he needed another data set to determine if his findings could be generalized. So McNorgan chose a math study, which relied on a mental multiplication task, and measured functional connectivity from the fMRI information in that second data set.

    advertisement

    Functional connectivity, unlike what the name might imply, is a dynamic description of how the brain is virtually wired from moment to moment. Don’t think in terms of the physical wires used in a network, but instead of how those wires are used throughout the day. When you’re working, your laptop is sending a document to your printer. Later in the day, your laptop might be streaming a movie to your television. How those wires are used depends on whether you’re working or relaxing. Functional connectivity changes according to the immediate task.
    The brain dynamically rewires itself according to the task all the time. Imagine reading a list of restaurant specials while standing only a few steps away from the menu board nailed to the wall. The visual cortex is working whenever you’re looking at something, but because you’re reading, the visual cortex works with, or is wired to, at least for the moment, the auditory cortex.
    Pointing to one of the items on the board, you accidentally knock it from the wall. When you reach out to catch it, your brain wiring changes. You’re no longer reading, but trying to catch a falling object, and your visual cortex now works with the pre-motor cortex to guide your hand.
    Different tasks, different wiring; or, as McNorgan explains, different functional networks.
    In the two data sets McNorgan used, participants were engaged in different tasks: language and math. Yet in each case, the connectivity fingerprint was the same, and he was able to identify dyslexia with 94% accuracy whether testing against the reading group or the math group.
    It was a whim, he said, to see how well his model distinguished good readers from poor readers — or from participants who weren’t reading at all. Seeing the accuracy, and the similarity, changed the direction of the paper McNorgan intended.
    Yes, he could identify dyslexia. But it became obvious that the brain’s wiring for reading was also present for math.
    Different task. Same functional networks.
    “The brain should be dynamically wiring itself in a way that’s specifically relevant to doing math because of the multiplication problem in the second data set, but there’s clear evidence of the dynamic configuration of the reading network showing up in the math task,” McNorgan says.
    He says it’s the sort of finding that strengthens the already strong case for supporting literacy.
    “These results show that the way our brain is wired for reading is actually influencing how the brain functions for math,” he said. “That says your reading skill is going to affect how you tackle problems in other domains, and helps us better understand children with learning difficulties in both reading and math.”
    As the line between cognitive domains becomes more blurred, McNorgan wonders what other domains the reading network is actually guiding.
    “I’ve looked at two domains which couldn’t be farther afield,” he said. “If the brain is showing that its wiring for reading is showing up in mental multiplication, what else might it be contributing toward?”
    That’s an open question, for now, according to McNorgan.
    “What I do know because of this research is that an educational emphasis on reading means much more than improving reading skill,” he said. “These findings suggest that learning how to read shapes so much more.” More

  • in

    Breakthrough lays groundwork for future quantum networks

    New Army-funded research could help lay the groundwork for future quantum communication networks and large-scale quantum computers.
    Researchers sent entangled qubit states through a communication cable linking one quantum network node to a second node.
    Scientists at the Pritzker School of Molecular Engineering at the University of Chicago, funded and managed by the U.S. Army Combat Capability Development, known as DEVCOM, Army Research Laboratory’s Center for Distributed Quantum Information, also amplified an entangled state via the same cable first by using the cable to entangle two qubits in each of two nodes, then entangling these qubits further with other qubits in the nodes. The peer-reviewed journal published the research in its Feb. 24, 2021, issue.
    “The entanglement distribution results the team achieved brought together years of their research related to approaches for transferring quantum states and related to advanced fabrication procedures to realize the experiments,” said Dr. Sara Gamble, program manager at the Army Research Office, an element of the Army’s corporate research laboratory, and co-manager of the CDQI, which funded the work. “This is an exciting achievement and one that paves the way for increasingly complex experiments with additional quantum nodes that we’ll need for the large-scale quantum networks and computers of ultimate interest to the Army.”
    Qubits, or quantum bits, are the basic units of quantum information. By exploiting their quantum properties, like superposition, and their ability to be entangled together, scientists and engineers are creating next-generation quantum computers that will be able solve previously unsolvable problems.
    The research team uses superconducting qubits, tiny cryogenic circuits that can be manipulated electrically.

    advertisement

    “Developing methods that allow us to transfer entangled states will be essential to scaling quantum computing,” said Prof. Andrew Cleland, the John A. MacLean senior professor of Molecular Engineering Innovation and Enterprise at University of Chicago, who led the research.
    Entanglement is a correlation that can be created between quantum entities such as qubits. When two qubits are entangled and a measurement is made on one, it will affect the outcome of a measurement made on the other, even if that second qubit is physically far away.
    Entanglement is a correlation that can be created between quantum entities such as qubits. When two qubits are entangled and a measurement is made on one, it will affect the outcome of a measurement made on the other, even if that second qubit is physically far away.
    To send the entangled states through the communication cable — a one-meter-long superconducting cable — the researchers created an experimental set-up with three superconducting qubits in each of two nodes. They connected one qubit in each node to the cable and then sent quantum states, in the form of microwave photons, through the cable with minimal loss of information. The fragile nature of quantum states makes this process quite challenging.
    The researchers developed a system in which the whole transfer process — node to cable to node — takes only a few tens of nanoseconds (a nanosecond is one billionth of a second). That allowed them to send entangled quantum states with very little information loss.

    advertisement

    The system also allowed them to amplify the entanglement of qubits. The researchers used one qubit in each node and entangled them together by essentially sending a half-photon through the cable. They then extended this entanglement to the other qubits in each node. When they were finished, all six qubits in two nodes were entangled in a single globally entangled state.
    “We want to show that superconducting qubits have a viable role going forward,” Cleland said.
    A quantum communication network could potentially take advantage of this advance. The group plans to extend their system to three nodes to build three-way entanglement.
    The researchers developed a system in which the whole transfer process — node to cable to node — takes only a few tens of nanoseconds (a nanosecond is one billionth of a second).
    “The team was able to identify a primary limiting factor in this current experiment related to loss in some of the components,” said Dr. Fredrik Fatemi, branch chief for quantum sciences, DEVCOM ARL, and co-manager of CDQI. “They have a clear path forward for increasingly complex experiments which will enable us to explore new regimes in distributed entanglement.” More

  • in

    Robots learn faster with quantum technology

    Artificial intelligence is part of our modern life. A crucial question for practical applications is how fast such intelligent machines can learn. An experiment has answered this question, showing that quantum technology enables a speed-up in the learning process. The physicists have achieved this result by using a quantum processor for single photons as a robot.
    Robots solving computer games, recognizing human voices, or helping in finding optimal medical treatments: those are only a few astonishing examples of what the field of artificial intelligence has produced in the past years. The ongoing race for better machines has led to the question of how and with what means improvements can be achieved. In parallel, huge recent progress in quantum technologies have confirmed the power of quantum physics, not only for its often peculiar and puzzling theories, but also for real-life applications. Hence, the idea of merging the two fields: on one hand, artificial intelligence with its autonomous machines; on the other hand, quantum physics with its powerful algorithms.
    Over the past few years, many scientists have started to investigate how to bridge these two worlds, and to study in what ways quantum mechanics can prove beneficial for learning robots, or vice versa. Several fascinating results have shown, for example, robots deciding faster on their next move, or the design of new quantum experiments using specific learning techniques. Yet, robots were still incapable of learning faster, a key feature in the development of increasingly complex autonomous machines.
    Within an international collaboration led by Philip Walther, a team of experimental physicists from the University of Vienna, together with theoreticians from the University of Innsbruck, the Austrian Academy of Sciences, the Leiden University, and the German Aerospace Center, have been successful in experimentally proving for the first time a speed-up in the actual robot’s learning time. The team has made use of single photons, the fundamental particles of light, coupled into an integrated photonic quantum processor, which was designed at the Massachusetts Institute of Technology. This processor was used as a robot and for implementing the learning tasks. Here, the robot would learn to route the single photons to a predefined direction. “The experiment could show that the learning time is significantly reduced compared to the case where no quantum physics is used,” says Valeria Saggio, first author of the publication.
    In a nutshell, the experiment can be understood by imagining a robot standing at a crossroad, provided with the task of learning to always take the left turn. The robot learns by obtaining a reward when doing the correct move. Now, if the robot is placed in our usual classical world, then it will try either a left or right turn, and will be rewarded only if the left turn is chosen. In contrast, when the robot exploits quantum technology, the bizarre aspects of quantum physics come into play. The robot can now make use of one of its most famous and peculiar features, the so called superposition principle. This can be intuitively understood by imagining the robot taking the two turns, left and right, at the same time. “This key feature enables the implementation of a quantum search algorithm that reduces the number of trials for learning the correct path. As a consequence, an agent that can explore its environment in superposition will learn significantly faster than its classical counterpart,” says Hans Briegel, who developed the theoretical ideas on quantum learning agents with his group at the University of Innsbruck.
    This experimental demonstration that machine learning can be enhanced by using quantum computing shows promising advantages when combining these two technologies. “We are just at the beginning of understanding the possibilities of quantum artificial intelligence” says Philip Walther, “and thus every new experimental result contributes to the development of this field, which is currently seen as one of the most fertile areas for quantum computing.”

    Story Source:
    Materials provided by University of Vienna. Note: Content may be edited for style and length. More

  • in

    Classic math conundrum solved: Superb algorithm for finding the shortest route

    One of the most classic algorithmic problems deals with calculating the shortest path between two points. A more complicated variant of the problem is when the route traverses a changing network — whether this be a road network or the internet. For 40 years, an algorithm has been sought to provide an optimal solution to this problem. Now, computer scientist Christian Wulff-Nilsen of the University of Copenhagen and two research colleagues have come up with a recipe.
    When heading somewhere new, most of us leave it to computer algorithms to help us find the best route, whether by using a car’s GPS, or public transport and map apps on their phone. Still, there are times when a proposed route doesn’t quite align with reality. This is because road networks, public transportation networks and other networks aren’t static. The best route can suddenly be the slowest, e.g. because a queue has formed due to roadworks or an accident.
    People probably don’t think about the complicated math behind routing suggestions in these types of situations. The software being used is trying to solve a variant for the classic algorithmic “shortest path” problem, the shortest path in a dynamic network. For 40 years, researchers have been working to find an algorithm that can optimally solve this mathematical conundrum. Now, Christian Wulff-Nilsen of the University of Copenhagen’s Department of Computer Science has succeeded in cracking the nut along with two colleagues.
    “We have developed an algorithm, for which we now have mathematical proof, that it is better than every other algorithm up to now — and the closest thing to optimal that will ever be, even if we look 1000 years into the future,” says Associate Professor Wulff-Nilsen. The results were presented at the FOCS 2020 conference.
    Optimally, in this context, refers to an algorithm that spends as little time and as little computer memory as possible to calculate the best route in a given network. This is not just true of road and transportation networks, but also the internet or any other type of network.
    Networks as graphs
    The researchers represent a network as a so-called dynamic graph.” In this context, a graph is an abstract representation of a network consisting of edges, roads for example, and nodes, representing intersections, for example. When a graph is dynamic, it means that it can change over time. The new algorithm handles changes consisting of deleted edges — for example, if the equivalent of a stretch of a road suddenly becomes inaccessible due to roadworks.

    advertisement

    “The tremendous advantage of seeing a network as an abstract graph is that it can be used to represent any type of network. It could be the internet, where you want to send data via as short a route as possible, a human brain or the network of friendship relations on Facebook. This makes graph algorithms applicable in a wide variety of contexts,” explains Christian Wulff-Nilsen.
    Traditional algorithms assume that a graph is static, which is rarely true in the real world. When these kinds of algorithms are used in a dynamic network, they need to be rerun every time a small change occurs in the graph — which wastes time.
    More data necessitates better algorithms
    Finding better algorithms is not just useful when travelling. It is necessary in virtually any area where data is produced, as Christian Wulff-Nilsen points out:
    “We are living in a time when volumes of data grow at a tremendous rate and the development of hardware simply can’t keep up. In order to manage all of the data we produce, we need to develop smarter software that requires less running time and memory. That’s why we need smarter algorithms,” he says.
    He hopes that it will be possible to use this algorithm or some of the techniques behind it in practice, but stresses that this is theoretical evidence and first requires experimentation.
    Background
    The research article “Near-Optimal Decremental SSSP in Dense Weighted Digraphs” was presented at the prestigious FOCS 2020 conference.
    The article was written by Christian Wulff-Nilsen, of the University of Copenhagen’s Department of Computer Science, and former Department of Computer Science PhD student Maximillian Probst Gutenberg and assistant professor Aaron Bernstein of Rutgers University.
    The version of the “shortest path” problem that the researchers solved is called “The Decremental Single-Source Shortest Path Problem.” It is essentially about maintaining the shortest paths in a changing dynamic network from one starting point to all other nodes in a graph. The changes to a network consist of edge removals.
    The paper gives a mathematical proof that the algorithm is essentially the optimal one for dynamic networks. On average, users will be able to change routes according to calculations made in constant time. More

  • in

    Large computer language models carry environmental, social risks

    Computer engineers at the world’s largest companies and universities are using machines to scan through tomes of written material. The goal? Teach these machines the gift of language. Do that, some even claim, and computers will be able to mimic the human brain.
    But this impressive compute capability comes with real costs, including perpetuating racism and causing significant environmental damage, according to a new paper, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” The paper is being presented Wednesday, March 10 at the ACM Conference on Fairness, Accountability and Transparency (ACM FAccT).
    This is the first exhaustive review of the literature surrounding the risks that come with rapid growth of language-learning technologies, said Emily M. Bender, a University of Washington professor of linguistics and a lead author of the paper along with Timnit Gebru, a well-known AI researcher.
    “The question we’re asking is what are the possible dangers of this approach and the answers that we’re giving involve surveying literature across a broad range of fields and pulling them together,” said Bender, who is the UW Howard and Frances Nostrand Endowed Professor.
    What the researchers surfaced was that there are downsides to the ever-growing computing power put into natural language models. They discuss how the ever-increasing size of training data for language modeling exacerbates social and environmental issues. Alarmingly, such language models perpetuate hegemonic language and can deceive people into thinking they are having a “real” conversation with a person rather than a machine. The increased computational needs of these models further contributes to environmental degradation.
    The authors were motivated to write the paper because of a trend within the field towards ever-larger language models and their growing spheres of influence.

    advertisement

    The paper already has generated wide-spread attention due, in part, to the fact that two of the paper’s co-authors say they were fired recently from Google for reasons that remain unsettled. Margaret Mitchell and Gebru, the two now-former Google researchers, said they stand by the paper’s scholarship and point to its conclusions as a clarion call to industry to take heed.
    “It’s very clear that putting in the concerns has to happen right now, because it’s already becoming too late,” said Mitchell, a researcher in AI.
    It takes an enormous amount of computing power to fuel the model language programs, Bender said. That takes up energy at tremendous scale, and that, the authors argue, causes environmental degradation. And those costs aren’t borne by the computer engineers, but rather by marginalized people who cannot afford the environmental costs.
    “It’s not just that there’s big energy impacts here, but also that the carbon impacts of that will bring costs first to people who are not benefiting from this technology,” Bender said. “When we do the cost-benefit analysis, it’s important to think of who’s getting the benefit and who’s paying the cost because they’re not the same people.”
    The large scale of this compute power also can restrict access to only the most well-resourced companies and research groups, leaving out smaller developers outside of the U.S., Canada, Europe and China. That’s because it takes huge machines to run the software necessary to make computers mimic human thought and speech.
    Another risk comes from the training data itself, the authors say. Because the computers read language from the Web and from other sources, they can pick up and perpetuate racist, sexist, ableist, extremist and other harmful ideologies.
    “One of the fallacies that people fall into is well, the internet is big, the internet is everything. If I just scrape the whole internet then clearly I’ve incorporated diverse viewpoints,” Bender said. “But when we did a step-by-step review of the literature, it says that’s not the case right now because not everybody’s on the internet, and of the people who are on the internet, not everybody is socially comfortable participating in the same way.”
    And, people can confuse the language models for real human interaction, believing that they’re actually talking with a person or reading something that a person has spoken or written, when, in fact, the language comes from a machine. Thus, the stochastic parrots.
    “It produces this seemingly coherent text, but it has no communicative intent. It has no idea what it’s saying. There’s no there there,” Bender said. More

  • in

    Robots can use eye contact to draw out reluctant participants in groups

    Eye contact is a key to establishing a connection, and teachers use it often to encourage participation. But can a robot do this too? Can it draw a response simply by making “eye” contact, even with people who are less inclined to speak up. A recent study suggests that it can.
    Researchers at KTH Royal Institute of Technology published results of experiments in which robots led a Swedish word game with individuals whose proficiency in the Nordic language was varied. They found that by redirecting its gaze to less proficient players, a robot can elicit involvement from even the most reluctant participants.
    Researchers Sarah Gillet and Ronald Cumbal say the results offer evidence that robots could play a productive role in educational settings.
    Calling on someone by name isn’t always the best way to elicit engagement, Gillet says. “Gaze can by nature influence very dynamically how much people are participating, especially if there is this natural tendency for imbalance — due to the differences in language proficiency,” she says.
    “If someone is not inclined to participate for some reason, we showed that gaze is able to overcome this difference and help everyone to participate.”
    Cumbal says that studies have shown that robots can support group discussion, but this is the first study to examine what happens when a robot uses gaze in a group interaction that isn’t balanced — when it is dominated by one or more individuals.
    The experiment involved pairs of players — one fluent in Swedish and one who is learning Swedish. The players were instructed to give the robot clues in Swedish so that it could guess the correct term. The face of the robot was an animated projection on a specially designed plastic mask.
    While it would be natural for a fluent speaker to dominate such a scenario, Cumbal says, the robot was able to prompt the participation of the less fluent player by redirecting its gaze naturally toward them and silently waiting for them to hazard an attempt.
    “Robot gaze can modify group dynamics — what role people take in a situation,” he says. “Our work builds on that and shows further that even when there is an imbalance in skills required for the activity, the gaze of a robot can still influence how the participants contribute.”

    Story Source:
    Materials provided by KTH, Royal Institute of Technology. Note: Content may be edited for style and length. More