Face recognition system ‘K-Eye’

Artificial intelligence (AI) is one of the key emerging technologies. Global IT companies are competitively launching the newest technologies and competition is heating up more than ever. However, most AI technologies focus on software and their operating speeds are low, making them a poor fit for mobile devices. Therefore, many big companies are investing to develop semiconductor chips for running AI programs with low power requirements but at high speeds.

A research team led by Professor Hoi-Jun Yoo of the Department of Electrical Engineering has developed a semiconductor chip, CNNP (CNN Processor), that runs AI algorithms with ultra-low power, and K-Eye, a face recognition system using CNNP. The system was made in collaboration with a start-up company, UX Factory Co.

The K-Eye series consists of two types: a wearable type and a dongle type. The wearable type device can be used with a smartphone via Bluetooth, and it can operate for more than 24 hours with its internal battery. Users hanging K-Eye around their necks can conveniently check information about people by using their smartphone or smart watch, which connects K-Eye and allows users to access a database via their smart devices. A smartphone with K-EyeQ, the dongle type device, can recognize and share information about users at any time.

When recognizing that an authorized user is looking at its screen, the smartphone automatically turns on without a passcode, fingerprint, or iris authentication. Since it can distinguish whether an input face is coming from a saved photograph versus a real person, the smartphone cannot be tricked by the user’s photograph.

The K-Eye series carries other distinct features. It can detect a face at first and then recognize it, and it is possible to maintain “Always-on” status with low power consumption of less than 1mW. To accomplish this, the research team proposed two key technologies: an image sensor with “Always-on” face detection and the CNNP face recognition chip.

The first key technology, the “Always-on” image sensor, can determine if there is a face in its camera range. Then, it can capture frames and set the device to operate only when a face exists, reducing the standby power significantly. The face detection sensor combines analog and digital processing to reduce power consumption. With this approach, the analog processor, combined with the CMOS Image Sensor array, distinguishes the background area from the area likely to include a face, and the digital processor then detects the face only in the selected area. Hence, it becomes effective in terms of frame capture, face detection processing, and memory usage.

The second key technology, CNNP, achieved incredibly low power consumption by optimizing a convolutional neural network (CNN) in the areas of circuitry, architecture, and algorithms. First, the on-chip memory integrated in CNNP is specially designed to enable data to be read in a vertical direction as well as in a horizontal direction. Second, it has immense computational power with 1024 multipliers and accumulators operating in parallel and is capable of directly transferring the temporal results to each other without accessing to the external memory or on-chip communication network. Third, convolution calculations with a two-dimensional filter in the CNN algorithm are approximated into two sequential calculations of one-dimensional filters to achieve higher speeds and lower power consumption.

With these new technologies, CNNP achieved 97% high accuracy but consumed only 1/5000 power of the GPU. Face recognition can be performed with only 0.62mW of power consumption, and the chip can show higher performance than the GPU by using more power.

These chips were developed by Kyeongryeol Bong, a Ph. D. student under Professor Yoo and presented at the International Solid-State Circuit Conference (ISSCC) held in San Francisco in February. CNNP, which has the lowest reported power consumption in the world, has achieved a great deal of attention and has led to the development of the present K-Eye series for face recognition.

Professor Yoo said “AI — processors will lead the era of the Fourth Industrial Revolution. With the development of this AI chip, we expect Korea to take the lead in global AI technology.”


Robot uses deep learning and big data to write and play its own music

A marimba-playing robot with four arms and eight sticks is writing and playing its own compositions in a lab at the Georgia Institute of Technology. The pieces are generated using artificial intelligence and deep learning.

Researchers fed the robot nearly 5,000 complete songs — from Beethoven to the Beatles to Lady Gaga to Miles Davis — and more than 2 million motifs, riffs and licks of music. Aside from giving the machine a seed, or the first four measures to use as a starting point, no humans are involved in either the composition or the performance of the music.

The first two compositions are roughly 30 seconds in length. The robot, named Shimon, can be seen and heard playing them at https://www.youtube.com/watch?v=j82nYLOnKtM and https://www.youtube.com/watch?v=6MSk5PP9KUA.

Ph.D. student Mason Bretan is the man behind the machine. He’s worked with Shimon for seven years, enabling it to “listen” to music played by humans and improvise over pre-composed chord progressions. Now Shimon is a solo composer for the first time, generating the melody and harmonic structure on its own.

“Once Shimon learns the four measures we provide, it creates its own sequence of concepts and composes its own piece,” said Bretan, who will receive his doctorate in music technology this summer at Georgia Tech. “Shimon’s compositions represent how music sounds and looks when a robot uses deep neural networks to learn everything it knows about music from millions of human-made segments.”

Bretan says this is the first time a robot has used deep learning to create music. And unlike its days of improvising, when it played monophonically, Shimon is able to play harmonies and chords. It’s also thinking much more like a human musician, focusing less on the next note, as it did before, and more on the overall structure of the composition.

“When we play or listen to music, we don’t think about the next note and only that next note,” said Bretan. “An artist has a bigger idea of what he or she is trying to achieve within the next few measures or later in the piece. Shimon is now coming up with higher-level musical semantics. Rather than thinking note by note, it has a larger idea of what it wants to play as a whole.”

Shimon was created by Bretan’s advisor, Gil Weinberg, director of Georgia Tech’s Center for Music Technology.

“This is a leap in Shimon’s musical quality because it’s using deep learning to create a more structured and coherent composition,” said Weinberg, a professor in the School of Music. “We want to explore whether robots could become musically creative and generate new music that we humans could find beautiful, inspiring and strange.”

Shimon will create more pieces in the future. As long as the researchers feed it a different seed, the robot will produce something different each time — music that the researchers can’t predict. In the first piece, Bretan fed Shimon a melody comprised of eighth notes. It received a sixteenth note melody the second time, which influenced it to generate faster note sequences.

Bretan acknowledges that he can’t pick out individual songs that Shimon is referencing. He is able to recognize classical chord progression and influences of artists, such as Mozart, for example. “They sound like a fusion of jazz and classical,” said Bretan, who plays the keyboards and guitar in his free time. “I definitely hear more classical, especially in the harmony. But then I hear chromatic moving steps in the first piece — that’s definitely something you hear in jazz.”

Shimon’s debut as a solo composer was featured in a video clip in the Consumer Electronic Show (CES) keynote and will have its first live performance at the Aspen Ideas Festival at the end of June. It’s the latest project within Weinberg’s lab. He and his students have also created a robotic prosthesis for a drummer, a robotic third arm for all drummers, and an interactive robotic companion that plays music from a phone and dances to the beat.

Story Source:

Materials provided by Georgia Institute of Technology. Original written by Jason Maderer. Note: Content may be edited for style and length.


Learning with light: New system allows optical ‘deep learning’

“Deep Learning” computer systems, based on artificial neural networks that mimic the way the brain learns from an accumulation of examples, have become a hot topic in computer science. In addition to enabling technologies such as face- and voice-recognition software, these systems could scour vast amounts of medical data to find patterns that could be useful diagnostically, or scan chemical formulas for possible new pharmaceuticals.

But the computations these systems must carry out are highly complex and demanding, even for the most powerful computers.

Now, a team of researchers at MIT and elsewhere has developed a new approach to such computations, using light instead of electricity, which they say could vastly improve the speed and efficiency of certain deep learning computations. Their results appear today in the journal Nature Photonics in a paper by MIT postdoc Yichen Shen, graduate student Nicholas Harris, professors Marin Soljacic and Dirk Englund, and eight others.

Soljacic says that many researchers over the years have made claims about optics-based computers, but that “people dramatically over-promised, and it backfired.” While many proposed uses of such photonic computers turned out not to be practical, a light-based neural-network system developed by this team “may be applicable for deep-learning for some applications,” he says.

Traditional computer architectures are not very efficient when it comes to the kinds of calculations needed for certain important neural-network tasks. Such tasks typically involve repeated multiplications of matrices, which can be very computationally intensive in conventional CPU or GPU chips.

After years of research, the MIT team has come up with a way of performing these operations optically instead. “This chip, once you tune it, can carry out matrix multiplication with, in principle, zero energy, almost instantly,” Soljacic says. “We’ve demonstrated the crucial building blocks but not yet the full system.”

By way of analogy, Soljacic points out that even an ordinary eyeglass lens carries out a complex calculation (the so-called Fourier transform) on the light waves that pass through it. The way light beams carry out computations in the new photonic chips is far more general but has a similar underlying principle. The new approach uses multiple light beams directed in such a way that their waves interact with each other, producing interference patterns that convey the result of the intended operation. The resulting device is something the researchers call a programmable nanophotonic processor.

The result, Shen says, is that the optical chips using this architecture could, in principle, carry out calculations performed in typical artificial intelligence algorithms much faster and using less than one-thousandth as much energy per operation as conventional electronic chips. “The natural advantage of using light to do matrix multiplication plays a big part in the speed up and power savings, because dense matrix multiplications are the most power hungry and time consuming part in AI algorithms” he says.

The new programmable nanophotonic processor, which was developed in the Englund lab by Harris and collaborators, uses an array of waveguides that are interconnected in a way that can be modified as needed, programming that set of beams for a specific computation. “You can program in any matrix operation,” Harris says. The processor guides light through a series of coupled photonic waveguides. The team’s full proposal calls for interleaved layers of devices that apply an operation called a nonlinear activation function, in analogy with the operation of neurons in the brain.

To demonstrate the concept, the team set the programmable nanophotonic processor to implement a neural network that recognizes four basic vowel sounds. Even with this rudimentary system, they were able to achieve a 77 percent accuracy level, compared to about 90 percent for conventional systems. There are “no substantial obstacles” to scaling up the system for greater accuracy, Soljacic says.

Englund adds that the programmable nanophotonic processor could have other applications as well, including signal processing for data transmission. “High-speed analog signal processing is something this could manage” faster than other approaches that first convert the signal to digital form, since light is an inherently analog medium. “This approach could do processing directly in the analog domain,” he says.

The team says it will still take a lot more effort and time to make this system useful; however, once the system is scaled up and fully functioning, it can find many user cases, such as data centers or security systems. The system could also be a boon for self-driving cars or drones, says Harris, or “whenever you need to do a lot of computation but you don’t have a lot of power or time.”

Comparing student performance on paper-and-pencil and computer-based-tests

Based on a study of more than 30,000 elementary, middle, and high school students conducted in winter 2015-16, researchers found that elementary and middle school students scored lower on a computer-based test that did not allow them to return to previous items than on two comparable tests — paper- or computer-based — that allowed them to skip, review, and change previous responses.

Elementary school students scored marginally higher on the computer-based exam that allowed them to go back to previous answers than on the paper-based exam, while there was no significant difference for middle school students on those two types of tests.

In contrast, high school students showed no difference in their performance on the three types of tests. Likewise, previous research has found that the option to skip, review, and change previous responses also had no effect on the test results of college students.

For the study, tests were given to students in grades 4-12 that assessed their understanding of energy through three testing systems. Instructors elected to administer either the paper-and-pencil test (PPT) or one of two computer-based tests (CBT) based on the availability of computers in their classrooms.

One CBT (using TAO, an open source online testing system) allowed students to skip items and freely move through the test, while the other CBT (using the AAAS assessment website) did not allow students to return to previous test items. In addition, on the TAO test, answers were selected by directly clicking on the text corresponding to an answer. On the AAAS exam, answers were chosen more indirectly, by clicking on a letter (A, B, C, or D) at the bottom of the screen corresponding with an answer.

Gender was found to have little influence on a student’s performance on PPT or CBT; however, students whose primary language was not English had lower performances on both CBTs compared to the PPT. The cause for the difference depending on primary language was unclear, but could have been linguistic challenges that the online environment presented or limits on opportunities to use computers in non-English-speaking environments.

Overall, the study results, along with previous research, indicate that being able to skip, review, and change previous responses could be beneficial for younger children in elementary and middle school but have no influence on older students in high school and college.

Furthermore, results indicated that marking an answer in a different location on a multiple-choice test could be challenging for younger students, students with poor organizational skills, students who have difficulties with concentration, or students who are physically impaired. In addition, having to match an answer to a corresponding letter at the bottom of the screen likely adds an additional level of complexity and cognitive processing.

The researchers note that additional study of CBT answer-choice selection and test navigation features and how they influence elementary and middle school students’ test performance is warranted.

The study was supported by a grant from the Institute of Education Sciences.


Smiling during victory could hurt future chances of cooperation

Smile and the whole world smiles with you? Well, not necessarily.

In a winning scenario, smiling can decrease your odds of success against the same opponent in subsequent matches, according to new research presented by the USC Institute for Creative Technologies and sponsored by the U.S. Army Research Laboratory.

People who smiled during victory increased the odds of their opponent acting aggressively to steal a pot of money rather than share it in future gameplay, according to a paper presented in May at the International Conference on Autonomous Agents and Multiagent Systems by USC ICT research assistant Rens Hoegen, USC ICT research programmer Giota Stratou and Jonathan Gratch, director of virtual humans research at USC ICT and a professor of computer science at the USC Viterbi School of Engineering.

Conversely, researchers found smiling during a loss tended to help the odds of success in the game going forward.

The study is in line with previous research published by senior author Gratch, whose main interest lies both in how people express these tells — an unconscious action that betrays deception — and using this data to create artificial intelligence to discern and even express these same emotional cues as a person.

“We think that emotion is the enemy of reason. But the truth is that emotion is our way of assigning value to things,” said Gratch. “Without it, we’d be faced with limitless choices.”

Gratch and other ICT researchers hope to imbue virtual humans and even robots with value-based assessment using emotional pattern recognition and reaction to form what might be called intuition or gut level decision-making.

Grin and bear it, but don’t gloat

Part of this research is accounting for the kind of emotion-based reasoning that might lead someone to act against their rational self-interest for the short-term satisfaction of “payback” — that is, cutting off their nose to spite their opponent’s smiling face.

For the AAMAS study, 370 participants played a version of the British television game show Golden Balls, where participants decide to “split” or “steal” a pot of money. If both participants choose “split,” they do just that — split the pot. If one player chooses to split with the other stealing, the latter gets the whole thing. If both choose to steal, neither wins.

Each participant was paid $30, with participants receiving additional tickets for a $100 lottery generated by their total number of successful “steals” and “splits.”

As participants played the game against each other on video Skype, reactions were recorded and encoded using emotion-tracking software that captures muscle movements in the face including cheek, lip and chin raises, dimples, and the compression and separation of lips.

As for the motivations of the players, researchers hypothesize that successful, smiling stealers open themselves to future punishment by the loser, while smiling during such a loss is seen as a gesture toward cooperation and a feeling of mutual success.

Teaching machines the power of a smile

In a similar study Gratch co-authored with ICT senior research associate Gale Lucas and colleagues in 2016, participants were shown to often misread honesty when negotiating with each other because reassuring cues like head movement, positive language and even smiling signal honesty, but actually more frequently represent dishonest action and behaviors.

Gratch has worked closely with the USC Marshall School of Business over the last several years to incorporate virtual humans that can understand these types of nuances into the study of negotiation. The Institute for Creative Technologies also works with agencies like the U.S. Army to use virtual humans in negotiation scenarios.

From Arthur Samuel’s checkers-playing AI of the 1950s and 1960s to the Joshua computer’s tic-tac-toe game of mutually assured destruction in the 1983 movie WarGames, artificial intelligence has been depicted as especially well-suited to beating people at their own, somewhat linear and strategy-based games.

IBM’s Deep Blue also famously and successfully battled chess master Garry Kasparov in the 1990s, and the computer system Watson did the same with its human opponents on Jeopardy! in 2011.

In the last year alone, different AIs have beaten top players in both the ancient game of Go and professional poker, the latter relying on bluffing, tells and accurate emotional readings of the opponent.


How the brain recognizes what the eye sees

If you think self-driving cars can’t get here soon enough, you’re not alone. But programming computers to recognize objects is very technically challenging, especially since scientists don’t fully understand how our own brains do it.

Now, Salk Institute researchers have analyzed how neurons in a critical part of the brain, called V2, respond to natural scenes, providing a better understanding of vision processing. The work is described in Nature Communications on June 8, 2017.

“Understanding how the brain recognizes visual objects is important not only for the sake of vision, but also because it provides a window on how the brain works in general,” says Tatyana Sharpee, an associate professor in Salk’s Computational Neurobiology Laboratory and senior author of the paper. “Much of our brain is composed of a repeated computational unit, called a cortical column. In vision especially we can control inputs to the brain with exquisite precision, which makes it possible to quantitatively analyze how signals are transformed in the brain.”

Although we often take the ability to see for granted, this ability derives from sets of complex mathematical transformations that we are not yet able to reproduce in a computer, according to Sharpee. In fact, more than a third of our brain is devoted exclusively to the task of parsing visual scenes.

Our visual perception starts in the eye with light and dark pixels. These signals are sent to the back of the brain to an area called V1 where they are transformed to correspond to edges in the visual scenes. Somehow, as a result of several subsequent transformations of this information, we then can recognize faces, cars and other objects and whether they are moving. How precisely this recognition happens is still a mystery, in part because neurons that encode objects respond in complicated ways.

Now, Sharpee and Ryan Rowekamp, a postdoctoral research associate in Sharpee’s group, have developed a statistical method that takes these complex responses and describes them in interpretable ways, which could be used to help decode vision for computer-simulated vision. To develop their model, the team used publicly available data showing brain responses of primates watching movies of natural scenes (such as forest landscapes) from the Collaborative Research in Computational Neuroscience (CRCNS) database.

“We applied our new statistical technique in order to figure out what features in the movie were causing V2 neurons to change their responses,” says Rowekamp. “Interestingly, we found that V2 neurons were responding to combinations of edges.”

The team revealed that V2 neurons process visual information according to three principles: first, they combine edges that have similar orientations, increasing robustness of perception to small changes in the position of curves that form object boundaries. Second, if a neuron is activated by an edge of a particular orientation and position, then the orientation 90 degrees from that will be suppressive at the same location, a combination termed “cross-orientation suppression.” These cross-oriented edge combinations are assembled in various ways to allow us to detect various visual shapes. The team found that cross-orientation was essential for accurate shape detection. The third principle is that relevant patterns are repeated in space in ways that can help perceive textured surfaces of trees or water and boundaries between them, as in impressionist paintings.

The researchers incorporated the three organizing principles into a model they named the Quadratic Convolutional model, which can be applied to other sets of experimental data. Visual processing is likely to be similar to how the brain processes smells, touch or sounds, the researchers say, so the work could elucidate processing of data from these areas as well.

“Models I had worked on before this weren’t entirely compatible with the data, or weren’t cleanly compatible,” says Rowekamp. “So it was really satisfying when the idea of combining edge recognition with sensitivity to texture started to pay off as a tool to analyze and understand complex visual data.”

But the more immediate application might be to improve object-recognition algorithms for self-driving cars or other robotic devices. “It seems that every time we add elements of computation that are found in the brain to computer-vision algorithms, their performance improves,” says Sharpee.

Story Source:

Materials provided by Salk Institute. Note: Content may be edited for style and length.


Self-learning robot hands

Researchers at Bielefeld University have developed a grasp system with robot hands that autonomously familiarizes itself with novel objects. The new system works without previously knowing the characteristics of objects, such as pieces of fruit or tools. It was developed as part of the large-scale research project Famula at Bielefeld University’s Cluster of Excellence Cognitive Interaction Technology (CITEC). The knowledge gained from this project could contribute to future service robots, for instance, that are able to independently adapt to working in new households. CITEC has invested approximately one million Euro in Famula. In a new “research_tv” report from Bielefeld University, the coordinators of the Famula project explain the new innovation.

“Our system learns by trying out and exploring on its own — just as babies approach new objects,” says neuroinformatics Professor Dr. Helge Ritter, who heads the Famula project together with sports scientist and cognitive psychologist Professor Dr. Thomas Schack and robotics Privatdozent Dr. Sven Wachsmuth.

The CITEC researchers are working on a robot with two hands that are based on human hands in terms of both shape and mobility. The robot brain for these hands has to learn how everyday objects like pieces of fruit, dishes, or stuffed animals can be distinguished on the basis of their color or shape, as well as what matters when attempting to grasp the object.

The Human Being as the Model

A banana can be held, and a button can be pressed. “The system learns to recognize such possibilities as characteristics, and constructs a model for interacting and re-identifying the object,” explains Ritter.

To accomplish this, the interdisciplinary project brings together work in artificial intelligence with research from other disciplines. Thomas Schack’s research group, for instance, investigated which characteristics study participants perceived to be significant in grasping actions. In one study, test subjects had to compare the similarity of more than 100 objects. “It was surprising that weight hardly plays a role. We humans rely mostly on shape and size when we differentiate objects,” says Thomas Schack. In another study, test subjects’ eyes were covered and they had to handle cubes that differed in weight, shape, and size. Infrared cameras recorded their hand movements. “Through this, we find out how people touch an object, and which strategies they prefer to use to identify its characteristics,” explains Dirk Koester, who is a member of Schack’s research group. “Of course, we also find out which mistakes people make when blindly handling objects.”

System Puts Itself in the Position of Its “Mentor”

Dr. Robert Haschke, a colleague of Helge Ritter, stands in front of a large metal cage with both robot arms and a table with various test objects. In his role as a human learning mentor, Dr. Haschke helps the system to acquire familiarity with novel objects, telling the robot hands which object on the table they should inspect next. To do this, Haschke points to individual objects, or gives spoken hints, such as in which direction an interesting object for the robot can be found (e.g. “behind, at left”). Using color cameras and depth sensors, two monitors display how the system perceives its surroundings and reacts to instructions from humans.

“In order to understand which objects they should work with, the robot hands have to be able to interpret not only spoken language, but also gestures,” explains Sven Wachsmuth, of CITEC’s Central Labs. “And they also have to be able to put themselves in the position of a human to also ask themselves if they have correctly understood.” Wachsmuth and his team are not only responsible for the system’s language capabilities: they have also given the system a face. From one of the monitors, Flobi follows the movements of the hands and reacts to the researchers’ instructions. Flobi is a stylized robot head that complements the robot’s language and actions with facial expressions. As part of the Famula system, the virtual version of the robot Flobi is currently in use.

Understanding Human Interaction

With the Famula project, CITEC researchers are conducting basic research that can benefit self-learning robots of the future in both the home and industry. “We want to literally understand how we learn to ‘grasp’ our environment with our hands. The robot makes it possible for us to test our findings in reality and to rigorously expose the gaps in our understanding. In doing so, we are contributing to the future use of complex, multi-fingered robot hands, which today are still too costly or complex to be used, for instance, in industry,” explains Ritter.

The project name Famula stands for “Deep Familiarization and Learning Grounded in Cooperative Manual Action and Language: From Analysis to Implementation.” The project has been running since 2014 and is limited to October 2017 at the moment. Eight research groups from the Cluster of Excellence CITEC are working together on the project. Famula is one of four large-scale projects at CITEC; other projects include a robot service apartment, the walking robot Hector, and the virtual coaching space ICSpace. As part of the Excellence Initiative of the Deutsche Forschungsgemeinschaft (German Research Foundation, DFG), CITEC is funded by state and federal governments (EXC 277).

Combatting weeds with lasers

A robot automatically identifies weeds in a field and combats them with a short laser pulse. Sustainable agriculture, which avoids the use of herbicides as far as possible, could benefit from this smart idea. Dr. Julio Pastrana and Tim Wigbels from the Institute of Geodesy and Geoinformation at the University of Bonn are convinced of this. With an EXIST Business Start-up Grant from the Federal Ministry for Economic Affairs and Energy, the scientists are now driving forward the development of this practical tool for field work.

Those who want a rich harvest need to drive back weeds so that the crops can grow better. In organic agriculture, herbicides are ruled out as they are considered toxic chemicals, and unwanted plants must be laboriously weeded out. If the expectations of Dr. Julio Pastrana and Tim Wigbels are correct, this time-consuming work can soon be taken care of by robots.

Laser-based weed control can eliminate herbicides

The computer scientists in the Photogrammetry Laboratory at the Institute for Geodesy and Geoinformation at the University of Bonn are currently developing a novel system: using cameras on an all-terrain robot vehicle or even a tractor add-on, unwanted wild weeds should be automatically identified in the various crops and combatted in a targeted way. “The robot shoots the leaves of the unwanted plants with short laser pulses, which causes a weakening in their vitality,” reports Dr. Pastrana. “It is thus predicted that we will no longer need to use herbicides on our fields and the environment will be protected,” adds Wigbels.

Before forming the start-up, Dr. Pastrana worked in robotics and researched automated image interpretation techniques with Prof. Cyrill Stachniss from the Institute of Geodesy and Geoinformation at the University of Bonn. Dr. Pastrana completed his doctorate on the detection and classification of weeds with the aid of statistical models at Leibniz Universität Hannover and built an earlier version of the robot there with a colleague. Wigbels studied Computer Engineering at RWTH Aachen University and then worked in software development within a company.

The researchers are now pushing forward their start-up “Escarda Technologies” for one year at the University of Bonn with an EXIST grant from the Federal Ministry for Economic Affairs and Energy. “It is now a case of finding investors and further developing the business plan for the start-up,” says Wigbels. The researchers are also using the funding from the Ministry to buy the parts needed to construct a prototype.

Prof. Stachniss is supporting the start-up in various ways: Pastrana and Wigbels can thus use laboratories at the institution and consult with colleagues there. What’s more, Rüdiger Wolf from Technology Transfer at the University of Bonn helped the start-up to submit the application for the EXIST funding. “The advice was very helpful,” says Dr. Pastrana, delighted. Both scientists would also like to participate in the start-up round tables organized by Technology Transfer in order to benefit from the experience of other start-ups. The EXIST grant also enables them to attend training programs to prepare them for the challenges of independence.

“The idea combines innovative robots with a current sustainability topic,” says transfer advisor Rüdiger Wolf. He says the analyses of the market and competition for such an application are sound. Pastrana is convinced of the benefits of the laser-based technique for new agricultural machinery: “Our aim is to contribute to achieving more sustainable agriculture.” At the Bonn Idea Exchange by the Bonn/Rhein-Sieg Chamber of Commerce and Industry, both founders won an award for the best start-up idea.

Story Source:

Materials provided by Universität Bonn. Note: Content may be edited for style and length.

Tactile sensor gives robots new capabilities

Eight years ago, Ted Adelson’s research group at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) unveiled a new sensor technology, called GelSight, that uses physical contact with an object to provide a remarkably detailed 3-D map of its surface.

Now, by mounting GelSight sensors on the grippers of robotic arms, two MIT teams have given robots greater sensitivity and dexterity. The researchers presented their work in two papers at the International Conference on Robotics and Automation last week.

In one paper, Adelson’s group uses the data from the GelSight sensor to enable a robot to judge the hardness of surfaces it touches — a crucial ability if household robots are to handle everyday objects.

In the other, Russ Tedrake’s Robot Locomotion Group at CSAIL uses GelSight sensors to enable a robot to manipulate smaller objects than was previously possible.

The GelSight sensor is, in some ways, a low-tech solution to a difficult problem. It consists of a block of transparent rubber — the “gel” of its name — one face of which is coated with metallic paint. When the paint-coated face is pressed against an object, it conforms to the object’s shape.

The metallic paint makes the object’s surface reflective, so its geometry becomes much easier for computer vision algorithms to infer. Mounted on the sensor opposite the paint-coated face of the rubber block are three colored lights and a single camera.

“[The system] has colored lights at different angles, and then it has this reflective material, and by looking at the colors, the computer … can figure out the 3-D shape of what that thing is,” explains Adelson, the John and Dorothy Wilson Professor of Vision Science in the Department of Brain and Cognitive Sciences.

In both sets of experiments, a GelSight sensor was mounted on one side of a robotic gripper, a device somewhat like the head of a pincer, but with flat gripping surfaces rather than pointed tips.

Contact points

For an autonomous robot, gauging objects’ softness or hardness is essential to deciding not only where and how hard to grasp them but how they will behave when moved, stacked, or laid on different surfaces. Tactile sensing could also aid robots in distinguishing objects that look similar.

In previous work, robots have attempted to assess objects’ hardness by laying them on a flat surface and gently poking them to see how much they give. But this is not the chief way in which humans gauge hardness. Rather, our judgments seem to be based on the degree to which the contact area between the object and our fingers changes as we press on it. Softer objects tend to flatten more, increasing the contact area.

The MIT researchers adopted the same approach. Wenzhen Yuan, a graduate student in mechanical engineering and first author on the paper from Adelson’s group, used confectionary molds to create 400 groups of silicone objects, with 16 objects per group. In each group, the objects had the same shapes but different degrees of hardness, which Yuan measured using a standard industrial scale.

Then she pressed a GelSight sensor against each object manually and recorded how the contact pattern changed over time, essentially producing a short movie for each object. To both standardize the data format and keep the size of the data manageable, she extracted five frames from each movie, evenly spaced in time, which described the deformation of the object that was pressed.

Finally, she fed the data to a neural network, which automatically looked for correlations between changes in contact patterns and hardness measurements. The resulting system takes frames of video as inputs and produces hardness scores with very high accuracy. Yuan also conducted a series of informal experiments in which human subjects palpated fruits and vegetables and ranked them according to hardness. In every instance, the GelSight-equipped robot arrived at the same rankings.

Yuan is joined on the paper by her two thesis advisors, Adelson and Mandayam Srinivasan, a senior research scientist in the Department of Mechanical Engineering; Chenzhuo Zhu, an undergraduate from Tsinghua University who visited Adelson’s group last summer; and Andrew Owens, who did his PhD in electrical engineering and computer science at MIT and is now a postdoc at the University of California at Berkeley.

Obstructed views

The paper from the Robot Locomotion Group was born of the group’s experience with the Defense Advanced Research Projects Agency’s Robotics Challenge (DRC), in which academic and industry teams competed to develop control systems that would guide a humanoid robot through a series of tasks related to a hypothetical emergency.

Typically, an autonomous robot will use some kind of computer vision system to guide its manipulation of objects in its environment. Such systems can provide very reliable information about an object’s location — until the robot picks the object up. Especially if the object is small, much of it will be occluded by the robot’s gripper, making location estimation much harder. Thus, at exactly the point at which the robot needs to know the object’s location precisely, its estimate becomes unreliable. This was the problem the MIT team faced during the DRC, when their robot had to pick up and turn on a power drill.

“You can see in our video for the DRC that we spend two or three minutes turning on the drill,” says Greg Izatt, a graduate student in electrical engineering and computer science and first author on the new paper. “It would be so much nicer if we had a live-updating, accurate estimate of where that drill was and where our hands were relative to it.”

That’s why the Robot Locomotion Group turned to GelSight. Izatt and his co-authors — Tedrake, the Toyota Professor of Electrical Engineering and Computer Science, Aeronautics and Astronautics, and Mechanical Engineering; Adelson; and Geronimo Mirano, another graduate student in Tedrake’s group — designed control algorithms that use a computer vision system to guide the robot’s gripper toward a tool and then turn location estimation over to a GelSight sensor once the robot has the tool in hand.

In general, the challenge with such an approach is reconciling the data produced by a vision system with data produced by a tactile sensor. But GelSight is itself camera-based, so its data output is much easier to integrate with visual data than the data from other tactile sensors.

In Izatt’s experiments, a robot with a GelSight-equipped gripper had to grasp a small screwdriver, remove it from a holster, and return it. Of course, the data from the GelSight sensor don’t describe the whole screwdriver, just a small patch of it. But Izatt found that, as long as the vision system’s estimate of the screwdriver’s initial position was accurate to within a few centimeters, his algorithms could deduce which part of the screwdriver the GelSight sensor was touching and thus determine the screwdriver’s position in the robot’s hand.


Meet the most nimble-fingered robot ever built

DexNet 2.0.

Credit: Adriel Olmos, CITRIS Media

Grabbing the awkwardly shaped items that people pick up in their day-to-day lives is a slippery task for robots. Irregularly shaped items such as shoes, spray bottles, open boxes, even rubber duckies are easy for people to grab and pick up, but robots struggle with knowing where to apply a grip. In a significant step toward overcoming this problem, roboticists at UC Berkeley have a built a robot that can pick up and move unfamiliar, real-world objects with a 99 percent success rate.

Berkeley professor Ken Goldberg, postdoctoral researcher Jeff Mahler and the Laboratory for Automation Science and Engineering (AUTOLAB) created the robot, called DexNet 2.0. DexNet 2.0’s high grasping success rate means that this technology could soon be applied in industry, with the potential to revolutionize manufacturing and the supply chain.

DexNet 2.0 gained its highly accurate dexterity through a process called deep learning. The researchers built a vast database of three-dimensional shapes — 6.7 million data points in total — that a neural network uses to learn grasps that will pick up and move objects with irregular shapes. The neural network was then connected to a 3D sensor and a robotic arm. When an object is placed in front of DexNet 2.0, it quickly studies the shape and selects a grasp that will successfully pick up and move the object 99 percent of the time. DexNet 2.0 is also three times faster than its previous version.

DexNet 2.0 was featured as the cover story of the latest issues of MIT Technology Review, which called DexNet 2.0 “the most nimble-fingered robot yet.” The complete paper will be published in July.

Story Source:

Materials provided by University of California – Berkeley. Note: Content may be edited for style and length.

Cite This Page:

University of California – Berkeley. “Meet the most nimble-fingered robot ever built.” ScienceDaily. ScienceDaily, 1 June 2017. .

University of California – Berkeley. (2017, June 1). Meet the most nimble-fingered robot ever built. ScienceDaily. Retrieved June 2, 2017 from www.sciencedaily.com/releases/2017/06/170601192717.htm

University of California – Berkeley. “Meet the most nimble-fingered robot ever built.” ScienceDaily. www.sciencedaily.com/releases/2017/06/170601192717.htm (accessed June 2, 2017).


Scientists slash computations for deep learning

Rice University computer scientists have adapted a widely used technique for rapid data lookup to slash the amount of computation — and thus energy and time — required for deep learning, a computationally intense form of machine learning.

“This applies to any deep-learning architecture, and the technique scales sublinearly, which means that the larger the deep neural network to which this is applied, the more the savings in computations there will be,” said lead researcher Anshumali Shrivastava, an assistant professor of computer science at Rice.

The research will be presented in August at the KDD 2017 conference in Halifax, Nova Scotia. It addresses one of the biggest issues facing tech giants like Google, Facebook and Microsoft as they race to build, train and deploy massive deep-learning networks for a growing body of products as diverse as self-driving cars, language translators and intelligent replies to emails.

Shrivastava and Rice graduate student Ryan Spring have shown that techniques from “hashing,” a tried-and-true data-indexing method, can be adapted to dramatically reduce the computational overhead for deep learning. Hashing involves the use of smart hash functions that convert data into manageable small numbers called hashes. The hashes are stored in tables that work much like the index in a printed book.

“Our approach blends two techniques — a clever variant of locality-sensitive hashing and sparse backpropagation — to reduce computational requirements without significant loss of accuracy,” Spring said. “For example, in small-scale tests we found we could reduce computation by as much as 95 percent and still be within 1 percent of the accuracy obtained with standard approaches.”

The basic building block of a deep-learning network is an artificial neuron. Though originally conceived in the 1950s as models for the biological neurons in living brains, artificial neurons are just mathematical functions, equations that act upon an incoming piece of data and transform it into an output.

In machine learning, all neurons start the same, like blank slates, and become specialized as they are trained. During training, the network is “shown” vast volumes of data, and each neuron becomes a specialist at recognizing particular patterns in the data. At the lowest layer, neurons perform the simplest tasks. In a photo recognition application, for example, low-level neurons might recognize light from dark or the edges of objects. Output from these neurons is passed on to the neurons in the next layer of the network, which search for their own specialized patterns. Networks with even a few layers can learn to recognize faces, dogs, stop signs and school buses.

“Adding more neurons to a network layer increases its expressive power, and there’s no upper limit to how big we want our networks to be,” Shrivastava said. “Google is reportedly trying to train one with 137 billion neurons.” By contrast, he said, there are limits to the amount of computational power that can be brought to bear to train and deploy such networks.

“Most machine-learning algorithms in use today were developed 30-50 years ago,” he said. “They were not designed with computational complexity in mind. But with ‘big data,’ there are fundamental limits on resources like compute cycles, energy and memory. Our lab focuses on addressing those limitations.”

Spring said computation and energy savings from hashing will be even larger on massive deep networks.

“The savings increase with scale because we are exploiting the inherent sparsity in big data,” he said. “For instance, let’s say a deep net has a billion neurons. For any given input — like a picture of a dog — only a few of those will become excited. In data parlance, we refer to that as sparsity, and because of sparsity our method will save more as the network grows in size. So while we’ve shown a 95 percent savings with 1,000 neurons, the mathematics suggests we can save more than 99 percent with a billion neurons.”

A copy of the paper “Scalable and Sustainable Deep Learning via Randomized Hashing” is available at: https://arxiv.org/abs/1602.08194

Story Source:

Materials provided by Rice University. Note: Content may be edited for style and length.


Artificial intelligence predicts patient lifespans

A computer’s ability to predict a patient’s lifespan simply by looking at images of their organs is a step closer to becoming a reality, thanks to new research led by the University of Adelaide.

The research, now published in the Nature journal Scientific Reports, has implications for the early diagnosis of serious illness, and medical intervention.

Researchers from the University’s School of Public Health and School of Computer Science, along with Australian and international collaborators, used artificial intelligence to analyse the medical imaging of 48 patients’ chests. This computer-based analysis was able to predict which patients would die within five years, with 69% accuracy — comparable to ‘manual’ predictions by clinicians.

This is the first study of its kind using medical images and artificial intelligence.

“Predicting the future of a patient is useful because it may enable doctors to tailor treatments to the individual,” says lead author Dr Luke Oakden-Rayner, a radiologist and PhD student with the University of Adelaide’s School of Public Health.

“The accurate assessment of biological age and the prediction of a patient’s longevity has so far been limited by doctors’ inability to look inside the body and measure the health of each organ.

“Our research has investigated the use of ‘deep learning’, a technique where computer systems can learn how to understand and analyse images.

“Although for this study only a small sample of patients was used, our research suggests that the computer has learnt to recognise the complex imaging appearances of diseases, something that requires extensive training for human experts,” Dr Oakden-Rayner says.

While the researchers could not identify exactly what the computer system was seeing in the images to make its predictions, the most confident predictions were made for patients with severe chronic diseases such as emphysema and congestive heart failure.

“Instead of focusing on diagnosing diseases, the automated systems can predict medical outcomes in a way that doctors are not trained to do, by incorporating large volumes of data and detecting subtle patterns,” Dr Oakden-Rayner says.

“Our research opens new avenues for the application of artificial intelligence technology in medical image analysis, and could offer new hope for the early detection of serious illness, requiring specific medical interventions.”

The researchers hope to apply the same techniques to predict other important medical conditions, such as the onset of heart attacks.

The next stage of their research involves analysing tens of thousands of patient images.

Story Source:

Materials provided by University of Adelaide. Note: Content may be edited for style and length.


Interactive tool helps novices and experts make custom robots

A new interactive design tool developed by Carnegie Mellon University’s Robotics Institute enables both novices and experts to build customized legged or wheeled robots using 3D-printed components and off-the-shelf actuators.

Using a familiar drag-and-drop interface, individuals can choose from a library of components and place them into the design. The tool suggests components that are compatible with each other, offers potential placements of actuators and can automatically generate structural components to connect those actuators.

Once the design is complete, the tool provides a physical simulation environment to test the robot before fabricating it, enabling users to iteratively adjust the design to achieve a desired look or motion.

“The process of creating new robotic systems today is notoriously challenging, time-consuming and resource-intensive,” said Stelian Coros, assistant professor of robotics. “In the not-so-distant future, however, robots will be part of the fabric of daily life and more people — not just roboticists — will want to customize robots. This type of interactive design tool would make this possible for just about anybody.”

Today, robotics Ph.D. student Ruta Desai will present a report on the design tool she developed with Coros and master’s student Ye Yuan at the IEEE International Conference on Robotics and Automation (ICRA 2017) in Singapore.

Coros’ team designed a number of robots with the tool and verified its feasibility by fabricating two — a wheeled robot with a manipulator arm that can hold a pen for drawing, and a four-legged “puppy” robot that can walk forward or sideways.

“The system makes it easy to experiment with different body proportions and motor configurations, and see how these decisions affect the robot’s ability to do certain tasks,” said Desai. “For instance, we discovered in simulation that some of our preliminary designs for the puppy enabled it to only walk forward, not sideways. We corrected that for the final design. The motions of the robot we actually built matched the desired motion we demonstrated in simulation very well.”

The research team developed models of how actuators, off-the-shelf brackets and 3D-printable structural components can be combined to form complex robotic systems. The iterative design process enables users to experiment by changing the number and location of actuators and to adjust the physical dimensions of the robot. The tool includes an auto-completion feature that allows it to automatically generate assemblies of components by searching through possible arrangements.

“Our work aims to make robotics more accessible to casual users,” Coros said. “This is important because people who play an active role in creating robotic devices for their own use are more likely to have positive feelings and higher quality interactions with them. This could accelerate the adoption of robots in everyday life.”

Story Source:

Materials provided by Carnegie Mellon University. Note: Content may be edited for style and length.

Shedding light on how humans walk, with robots

Learning how to walk is difficult for toddlers to master; it’s even harder for adults who are recovering from a stroke, traumatic brain injury, or other condition, requiring months of intensive, often frustrating physical therapy. With the recent boom of the robotic exoskeleton industry, more and more patients are being strapped into machines that apply forces to their legs as they walk, gently prodding them to modify their movements by lengthening their strides, straightening their hips, and bending their knees. But, are all patients benefiting from this kind of treatment? A group of scientists led by Paolo Bonato, Ph.D., Associate Faculty member at the Wyss Institute for Biologically Inspired Engineering at Harvard University and Director of the Motion Analysis Laboratory at Spaulding Rehabilitation Hospital, has discovered a crucial caveat for rehabilitative exoskeletons: humans whose lower limbs are fastened to a typical clinical robot only modify their gait if the forces the robot applies threaten their walking stability.

In a study published in the newest issue of Science Robotics, the researchers measured how test subjects’ gait changed in response to forces applied by a robotic exoskeleton as they walked on a treadmill. To the team’s surprise, the walkers adjusted their stride in response to a change in the length, but not the height, of their step, even when step height and length were disturbed at the same time. The scientists believe that this discrepancy can be explained by the central nervous system (CNS)’s primary reliance on stability when determining how to adjust to a disruption in normal walking. “Lifting your foot higher mid-stride doesn’t really make you that much less stable, whereas placing your foot closer or further away from your center of mass can really throw off your balance, so the body adjusts much more readily to that disturbance,” says Giacomo Severini, Ph.D., one of the three first authors of the paper, who is now an Assistant Professor at University College Dublin.

In fact, the brain is so willing to adapt to instability that it will expend a significant amount of the body’s energy to do so, most likely because the consequences of wobbly walking can be severe: a broken ankle, torn ligaments, or even a fall from a height. However, this prioritization of stability means that other aspects of walking, like the height of the foot off the ground or the angle of the toes, may require treatment beyond walking in a clinical exoskeleton. “To modify step height, for example, you’d need to design forces so that the change in height, which the brain normally interprets as neutral, becomes challenging to the patient’s balance,” says Severini. Most robots used in clinical settings today do not allow for that kind of customization.

The brain appears to create an internal model of the body’s movement based on the environment and its normal gait, and effectively predicts each step. When reality differs from that model (i.e., when a force is applied), the brain adjusts the body’s step length accordingly to compensate until the force is removed and the body recalibrates to the mental model. “The results of our study give us insight into the way people adapt to external forces while walking in general, which is useful for clinicians when evaluating whether their patients will respond to clinical robot interventions,” says Bonato, who is also an Associate Professor at Harvard Medical School (HMS).

“The results of this research are very important from a clinical point of view,” agrees Ross Zafonte, D.O., Chairperson of the Department of Physical Medicine and Rehabilitation at HMS and Senior Vice President of Medical Affairs Research and Education at Spaulding Rehabilitation Hospital. “It is thanks to advances in our understanding of the interactions between robots and patients, such as the ones investigated in this study, that we can design effective robot-assisted gait therapy.”

“As the human population ages, robotics is playing an increasing role in their care and treatment,” says Donald Ingber, M.D., Ph.D., Founding Director of the Wyss Institute, who is also the Judah Folkman Professor of Vascular Biology at HMS and Boston Children’s Hospital, and Professor of Bioengineering at Harvard’s John A. Paulson School of Engineering and Applied Sciences (SEAS). “Studying how the human body interacts with robots can not only teach us how to build better clinical rehabilitation machines, but also how our own human bodies work.”

Story Source:

Materials provided by Wyss Institute for Biologically Inspired Engineering at Harvard. Original written by Lindsay Brownell. Note: Content may be edited for style and length.

Parasitic robot system for waypoint navigation of turtle

A KAIST research team presented a hybrid animal-robot interaction called “the parasitic robot system,” that imitates the nature relationship between parasites and host.

The research team led by Professor Phil-Seung Lee of the Department of Mechanical Engineering took an animal’s locomotive abilities to apply the theory of using a robot as a parasite. The robot is attached to its host animal in a way similar to an actual parasite, and it interacts with the host through particular devices and algorithms.

Even with remarkable technology advancements, robots that operate in complex and harsh environments still have some serious limitations in moving and recharging. However, millions of years of evolution have led to there being many real animals capable of excellent locomotion and survive in actual natural environment.

Certain kinds of real parasites can manipulate the behavior of the host to increase the probability of its own reproduction. Similarly, in the proposed concept of a “parasitic robot,” a specific behavior is induced by the parasitic robot in its host to benefit the robot.

The team chose a turtle as their first host animal and designed a parasitic robot that can perform “stimulus-response training.” The parasitic robot, which is attached to the turtle, can induce the turtle’s object-tracking behavior through repeated training sessions.

The robot then simply guides it using LEDs and feeds it snacks as a reward for going in the right direction through a programmed algorithm. After training sessions lasting five weeks, the parasitic robot can successfully control the direction of movement of the host turtles in the waypoint navigation task in a water tank.

This hybrid animal-robot interaction system could provide an alternative solution of the limitations of conventional mobile robot systems in various fields. Ph.D. candidate Dae-Gun Kim, the first author of this research said that there are a wide variety of animals including mice, birds, and fish that could perform equally as well at such tasks. He said that in the future, this system will be applied to various exploration and reconnaissance missions that humans and robots find it difficult to do on their own.

Kim said, “This hybrid animal-robot interaction system could provide an alternative solution to the limitations of conventional mobile robot systems in various fields, and could also act as a useful interaction system for the behavioral sciences.”

Story Source:

Materials provided by The Korea Advanced Institute of Science and Technology (KAIST). Note: Content may be edited for style and length.