Aurelia Butler

Computer science researchers have demonstrated that a widely used technique called neural network pruning can adversely affect the performance of deep learning models, detailed what causes these performance problems, and demonstrated a technique for addressing the challenge.
Deep learning is a type of artificial intelligence that can be used to classify things, such as images, text or sound. For example, it can be used to identify individuals based on facial images. However, deep learning models often require a lot of computing resources to operate. This poses challenges when a deep learning model is put into practice for some applications.
To address these challenges, some systems engage in “neural network pruning.” This effectively makes the deep learning model more compact and, therefore, able to operate while using fewer computing resources.
“However, our research shows that this network pruning can impair the ability of deep learning models to identify some groups,” says Jung-Eun Kim, co-author of a paper on the work and an assistant professor of computer science at North Carolina State University.
“For example, if a security system uses deep learning to scan people’s faces in order to determine whether they have access to a building, the deep learning model would have to be made compact so that it can operate efficiently. This may work fine most of the time, but the network pruning could also affect the deep learning model’s ability to identify some faces.”
In their new paper, the researchers lay out why network pruning can adversely affect the performance of the model at identifying certain groups — which the literature calls “minority groups” — and demonstrate a new technique for addressing these challenges. More

The future of work is here.
As industries begin to see humans working closely with robots, there’s a need to ensure that the relationship is effective, smooth and beneficial to humans. Robot trustworthiness and humans’ willingness to trust robot behavior are vital to this working relationship. However, capturing human trust levels can be difficult due to subjectivity, a challenge researchers in the Wm Michael Barnes ’64 Department of Industrial and Systems Engineering at Texas A&M University aim to solve.
Dr. Ranjana Mehta, associate professor and director of the NeuroErgonomics Lab, said her lab’s human-autonomy trust research stemmed from a series of projects on human-robot Interactions in safety-critical work domains funded by the National Science Foundation (NSF).
“While our focus so far was to understand how operator states of fatigue and stress impact how humans interact with robots, trust became an important construct to study,” Mehta said. “We found that as humans get tired, they let their guards down and become more trusting of automation than they should. However, why that is the case becomes an important question to address.”
Mehta’s latest NSF-funded work, recently published in Human Factors: The Journal of the Human Factors and Ergonomics Society, focuses on understanding the brain-behavior relationships of why and how an operator’s trusting behaviors are influenced by both human and robot factors.
Mehta also has another publication in the journal Applied Ergonomics that investigates these human and robot factors.
Using functional near-infrared spectroscopy, Mehta’s lab captured functional brain activity as operators collaborated with robots on a manufacturing task. They found faulty robot actions decreased the operator’s trust in the robots. That distrust was associated with increased activation of regions in the frontal, motor and visual cortices, indicating increasing workload and heightened situational awareness. Interestingly, the same distrusting behavior was associated with the decoupling of these brain regions working together, which otherwise were well connected when the robot behaved reliably. Mehta said this decoupling was greater at higher robot autonomy levels, indicating that neural signatures of trust are influenced by the dynamics of human-autonomy teaming.
“What we found most interesting was that the neural signatures differed when we compared brain activation data across reliability conditions (manipulated using normal and faulty robot behavior) versus operator’s trust levels (collected via surveys) in the robot,” Mehta said. “This emphasized the importance of understanding and measuring brain-behavior relationships of trust in human-robot collaborations since perceptions of trust alone is not indicative of how operators’ trusting behaviors shape up.”
Dr. Sarah Hopko ’19, lead author on both papers and recent industrial engineering doctoral student, said neural responses and perceptions of trust are both symptoms of trusting and distrusting behaviors and relay distinct information on how trust builds, breaches and repairs with different robot behaviors. She emphasized the strengths of multimodal trust metrics — neural activity, eye tracking, behavioral analysis, etc. — can reveal new perspectives that subjective responses alone cannot offer.
The next step is to expand the research into a different work context, such as emergency response, and understand how trust in multi-human robot teams impact teamwork and taskwork in safety-critical environments. Mehta said the long-term goal is not to replace humans with autonomous robots but to support them by developing trust-aware autonomy agents.
“This work is critical, and we are motivated to ensure that humans-in-the-loop robotics design, evaluation and integration into the workplace are supportive and empowering of human capabilities,” Mehta said.
Story Source:
Materials provided by Texas A&M University. Original written by Jennifer Reiley. Note: Content may be edited for style and length. More

Many computer systems people interact with on a daily basis require knowledge about certain aspects of the world, or models, to work. These systems have to be trained, often needing to learn to recognize objects from video or image data. This data often contains superfluous content that reduces the accuracy of models. So researchers found a way to incorporate natural hand gestures into the teaching process. This way, users can more easily teach machines about objects, and the machines can also learn more effectively.
You’ve probably heard the term machine learning before, but are you familiar with machine teaching? Machine learning is what happens behind the scenes when a computer uses input data to form models that can later be used to perform useful functions. But machine teaching is the somewhat less explored part of the process, of how the computer gets its input data to begin with. In the case of visual systems, for example ones that can recognize objects, people need to show objects to a computer so it can learn about them. But there are drawbacks to the ways this is typically done that researchers from the University of Tokyo’s Interactive Intelligent Systems Laboratory sought to improve.
“In a typical object training scenario, people can hold an object up to a camera and move it around so a computer can analyze it from all angles to build up a model,” said graduate student Zhongyi Zhou. “However, machines lack our evolved ability to isolate objects from their environments, so the models they make can inadvertently include unnecessary information from the backgrounds of the training images. This often means users must spend time refining the generated models, which can be a rather technical and time-consuming task. We thought there must be a better way of doing this that’s better for both users and computers, and with our new system, LookHere, I believe we have found it.”
Zhou, working with Associate Professor Koji Yatani, created LookHere to address two fundamental problems in machine teaching: firstly, the problem of teaching efficiency, aiming to minimize the users’ time, and required technical knowledge. And secondly, of learning efficiency — how to ensure better learning data for machines to create models from. LookHere achieves these by doing something novel and surprisingly intuitive. It incorporates the hand gestures of users into the way an image is processed before the machine incorporates it into its model, known as HuTics. For example, a user can point to or present an object to the camera in a way that emphasizes its significance compared to the other elements in the scene. This is exactly how people might show objects to each other. And by eliminating extraneous details, thanks to the added emphasis to what’s actually important in the image, the computer gains better input data for its models.
“The idea is quite straightforward, but the implementation was very challenging,” said Zhou. “Everyone is different and there is no standard set of hand gestures. So, we first collected 2,040 example videos of 170 people presenting objects to the camera into HuTics. These assets were annotated to mark what was part of the object and what parts of the image were just the person’s hands. LookHere was trained with HuTics, and when compared to other object recognition approaches, can better determine what parts of an incoming image should be used to build its models. To make sure it’s as accessible as possible, users can use their smartphones to work with LookHere and the actual processing is done on remote servers. We also released our source code and data set so that others can build upon it if they wish.”
Factoring in the reduced demand on users’ time that LookHere affords people, Zhou and Yatani found that it can build models up to 14 times faster than some existing systems. At present, LookHere deals with teaching machines about physical objects and it uses exclusively visual data for input. But in theory, the concept can be expanded to use other kinds of input data such as sound or scientific data. And models made from that data would benefit from similar improvements in accuracy too.
Story Source:
Materials provided by University of Tokyo. Note: Content may be edited for style and length. More

Quantum dots are clusters of some 1,000 atoms which act as one large ‘super-atom’. It is possible to accurately design the electronic properties of these dots just by changing their size. However, to create functional devices, a large number of dots have to be combined into a new material. During this process, the properties of the dots are often lost. Now, a team led by University of Groningen professor of Photophysics and Optoelectronics, Maria Antonietta Loi, has succeeded in making a highly conductive optoelectronic metamaterial through self-organization. The metamaterial is described in the journal Advanced Materials, published on 29 October.
Quantum dots of PbSe (lead selenide) or PbS (lead sulphide) can convert shortwave infrared light into an electrical current. This is a useful property for making detectors, or a switch for telecommunications. ‘However, a single dot does not make a device. And when dots are combined, the assembly often loses the unique optical properties of individual dots, or, if they do maintain them, their capacity to transport charge carriers becomes very poor’, explains Loi. ‘This is because it is difficult to create an ordered material from the dots.’
Ordered layer
Working with colleagues from the Zernike Institute for Advanced Materials at the Faculty of Science and Engineering, University of Groningen, Loi experimented with a method that allows the production of a metamaterial from a colloidal solution of quantum dots. These dots, each about five to six nanometres in size, show a very high conductivity when assembled in an ordered array, while maintaining their optical properties.
‘We knew from the literature that dots can self-organize into a two-dimensional, ordered layer. We wanted to expand this to a 3D material’, says Loi. To achieve this, they filled small containers with a liquid that acted as a ‘mattress’ for the colloidal quantum dots. ‘By injecting a small amount onto the surface of the liquid, we created a 2D material. Then, adding a bigger volume of quantum dots turned out to produce an ordered 3D material.’
Superlattice
The dots are not submersed in the liquid, but self-orient on the surface to achieve a low energy state. ‘The dots have a truncated cubic shape, and when they are put together, they form an ordered structure in three dimensions; a superlattice, where the dots act like atoms in a crystal’, explains Loi. This superlattice that is composed by the quantum dot super atoms displays the highest electron mobility reported for quantum dot assemblies.
Detectors
It took special equipment to ascertain what the new metamaterial looks like. The team used an electron microscope which has atomic resolution to show the details of the material. They also ‘imaged’ the large-scale structure of the material using a technique called Grazing-incidence small-angle X-ray scattering. ‘Both techniques are available at the Zernike Institute, thanks to my colleagues Bart Kooi and Giuseppe Portale, respectively, which was a great help’, says Loi.
Measurements of the electronic properties of the material show that it closely resembles that of a bulk semiconductor, but with the optical properties of the dots. Thus, the experiment paves the way to create new metamaterials based on quantum dots. The sensitivity of the dots used in the present study to infrared light could be used to create optical switches for telecommunication devices. ‘And they might also be used in infrared detectors for night-vision and autonomous driving.’
ERC Grant
Loi is extremely pleased with the results of the experiments: ‘People have been dreaming of achieving this since the 1980s. That is how long attempts have been made to assemble quantum dots into functional materials. The control of the structure and the properties we have achieved was beyond our wildest expectations.’ Loi is now working on understanding and improving the technology to build extended superlattices from quantum dots, but is also planning to do so with other building blocks, for which she was recently awarded an Advanced Grant from the European Research Council. ‘Our next step is to improve the technique in order to make the materials more perfect and fabricate photodetectors with them.’
Story Source:
Materials provided by University of Groningen. Note: Content may be edited for style and length. More

Engineers at MIT and the University of Tokyo have produced centimeter-scale structures, large enough for the eye to see, that are packed with hundreds of billions of hollow aligned fibers, or nanotubes, made from hexagonal boron nitride.
Hexagonal boron nitride, or hBN, is a single-atom-thin material that has been coined “white graphene” for its transparent appearance and its similarity to carbon-based graphene in molecular structure and strength. It can also withstand higher temperatures than graphene, and is electrically insulating, rather than conductive. When hBN is rolled into nanometer-scale tubes, or nanotubes, its exceptional properties are significantly enhanced.
The team’s results, published today in the journal ACS Nano, provide a route toward fabricating aligned boron nitride nanotubes (A-BNNTs) in bulk. The researchers plan to harness the technique to fabricate bulk-scale arrays of these nanotubes, which can then be combined with other materials to make stronger, more heat-resistant composites, for instance to shield space structures and hypersonic aircraft.
As hBN is transparent and electrically insulating, the team also envisions incorporating the BNNTs into transparent windows and using them to electrically insulate sensors within electronic devices. The team is also investigating ways to weave the nanofibers into membranes for water filtration and for “blue energy” — a concept for renewable energy in which electricity is produced from the ionic filtering of salt water into fresh water.
Brian Wardle, professor of aeronautics and astronautics at MIT, likens the team’s results to scientists’ decades-long, ongoing pursuit of manufacturing bulk-scale carbon nanotubes.
“In 1991, a single carbon nanotube was identified as an interesting thing, but it’s been 30 years getting to bulk aligned carbon nanotubes, and the world’s not even fully there yet,” Wardle says. “With the work we’re doing, we’ve just short-circuited about 20 years in getting to bulk-scale versions of aligned boron nitride nanotubes.”
Wardle is the senior author of the new study, which includes lead author and MIT research scientist Luiz Acauan, former MIT postdoc Haozhe Wang, and collaborators at the University of Tokyo. More

Stake out in Pennsylvania’s Cook State Forest at the right time of year and you can see one of nature’s great light shows: swarms of fireflies that synchronize their flashes like strings of Christmas lights in the dark.
A new study by Pitt mathematicians shows that math borrowed from neuroscience can describe how swarms of these unique insects coordinate their light show, capturing key details about how they behave in the wild.
“This firefly has a quick sequence of flashes, and then a big pause before the next burst,” said Jonathan Rubin, professor and chair of the Department of Mathematics in the Kenneth P. Dietrich School of Arts and Sciences. “We knew a good framework for modeling this that could capture a lot of the features, and we were curious how far we could push it.”
Male fireflies produce a glow from their abdomens to call out to potential mates, sending out blinking patterns in the dark to woo females of their own species. Synchronous fireflies of the species Photinus carolinus take it a step further, coordinating their blinking throughout entire swarms. It’s a rare trait — there are only a handful of such species in North America — and the striking lights they produce draw crowds to locations where the insects are known to gather.
They’ve also attracted the interest of mathematicians seeking to understand how they synchronize their blinks. It’s just one example of how synchronization can evolve from randomness, a process that has intrigued mathematicians for centuries. One famous example from the 1600s showed that pendulum clocks hung next to one another synchronize through vibrations that travel through the wall, and the same branch of math can be used to describe everything from the action of intestines to audience members clapping.
“Synchrony is important for a lot of things, good and bad,” said co-author Bard Ermentrout, distinguished professor of mathematics in the Dietrich School. “Physicists, mathematicians, we’re all interested in synchronization.”
To crack the fireflies’ light show, the Pitt team used a more complex model called an “elliptic burster” that’s used to describe the behavior of brain cells. The duo, along with then-undergrad Madeline McCrea (A&S ’22) published details of their model Oct. 26 in the Journal of the Royal Society Interface. More

Most deaths from melanoma — the most lethal form of skin cancer — occur in patients who were initially diagnosed with early-stage melanoma and then later experienced a recurrence that is typically not detected until it has spread or metastasized.
A team led by investigators at Massachusetts General Hospital (MGH) recently developed an artificial intelligence-based method to predict which patients are most likely to experience a recurrence and are therefore expected to benefit from aggressive treatment. The method was validated in a study published in npj Precision Oncology.
Most patients with early-stage melanoma are treated with surgery to remove cancerous cells, but patients with more advanced cancer often receive immune checkpoint inhibitors, which effectively strengthen the immune response against tumor cells but also carry significant side effects.
“There is an urgent need to develop predictive tools to assist in the selection of high-risk patients for whom the benefits of immune checkpoint inhibitors would justify the high rate of morbid and potentially fatal immunologic adverse events observed with this therapeutic class,” says senior author Yevgeniy R. Semenov, MD, an investigator in the Department of Dermatology at MGH.
“Reliable prediction of melanoma recurrence can enable more precise treatment selection for immunotherapy, reduce progression to metastatic disease and improve melanoma survival while minimizing exposure to treatment toxicities.”
To help achieve this, Semenov and his colleagues assessed the effectiveness of algorithms based on machine learning, a branch of artificial intelligence, that used data from patient electronic health records to predict melanoma recurrence.
Specifically, the team collected 1,720 early-stage melanomas — 1,172 from the Mass General Brigham healthcare system (MGB) and 548 from the Dana-Farber Cancer Institute (DFCI) — and extracted 36 clinical and pathologic features of these cancers from electronic health records to predict patients’ recurrence risk with machine learning algorithms. Algorithms were developed and validated with various MGB and DFCI patient sets, and tumor thickness and rate of cancer cell division were identified as the most predictive features.
“Our comprehensive risk prediction platform using novel machine learning approaches to determine the risk of early-stage melanoma recurrence reached high levels of classification and time to event prediction accuracy,” says Semenov. “Our results suggest that machine learning algorithms can extract predictive signals from clinicopathologic features for early-stage melanoma recurrence prediction, which will enable the identification of patients who may benefit from adjuvant immunotherapy.”
Additional Mass General co-authors include Ahmad Rajeh, Michael R. Collier, Min Seok Choi, Munachimso Amadife, Kimberly Tang, Shijia Zhang, Jordan Phillips, Nora A. Alexander, Yining Hua, Wenxin Chen, Diane, Ho, Stacey Duey, and Genevieve M. Boland.
This work was supported by the Melanoma Research Alliance, the National Institutes of Health, the Department of Defense, and the Dermatology Foundation.
Story Source:
Materials provided by Massachusetts General Hospital. Note: Content may be edited for style and length. More