Aurelia Butler

A linguistic bias in the English language that leads us to ‘improve’ things by adding to them, rather than taking away, is so common that it is even ingrained in AI chatbots, a new study reveals.
Language related to the concept of ‘improvement’ is more closely aligned with addition, rather than subtraction. This can lead us to make decisions which can overcomplicate things we are trying to make better.
The study is published today (Monday 3rd April) in Cognitive Science, by an international research team from the Universities of Birmingham, Glasgow, Potsdam, and Northumbria University.
Dr Bodo Winter, Associate Professor in Cognitive Linguistics at the University of Birmingham said: “Our study builds on existing research which has shown that when people seek to make improvements, they generally add things.
“We found that the same bias is deeply embedded in the English language. For example, the word ‘improve’ is closer in meaning to words like ‘add’ and ‘increase’ than to ‘subtract’ and ‘decrease’, so when somebody at a meeting says, ‘Does anybody have ideas for how we could improve this?,’ it will already, implicitly, contain a call for improving by adding rather than improving by subtracting.”
The research also finds that other verbs of change like ‘to change’, ‘to modify’, ‘to revise’ or ‘to enhance’ behave in a similar way, and if this linguistic addition bias is left unchecked, it can make things worse, rather than improve them. For example, improving by adding rather than subtracting can make bureaucracy become excessive.
This bias works in reverse as well. Addition-related words are more frequent and more positive in ‘improvement’ contexts rather than subtraction-related words, meaning this addition bias is found at multiple levels of English language structure and use.
The bias is so ingrained that even AI chatbots have it built in. The researchers asked GPT-3, the predecessor of ChatGPT, what it thought of the word ‘add’. It replied: “The word ‘add’ is a positive word. Adding something to something else usually makes it better. For example, if you add sugar to your coffee, it will probably taste better. If you add a new friend to your life, you will probably be happier.”
Dr Winter concludes: “The positive addition bias in the English language is something we should all be aware of. It can influence our decisions and mean we are pre-disposed to add more layers, more levels, more things when in fact we might actually benefit from removing or simplifying.
“Maybe next time we are asked at work, or in life, to come up with suggestions on how to make improvements, we should take a second to consider our choices for a bit longer.” More

The cosmos would look a lot better if Earth’s atmosphere wasn’t photo bombing it all the time.
Even images obtained by the world’s best ground-based telescopes are blurry due to the atmosphere’s shifting pockets of air. While seemingly harmless, this blur obscures the shapes of objects in astronomical images, sometimes leading to error-filled physical measurements that are essential for understanding the nature of our universe.
Now researchers at Northwestern University and Tsinghua University in Beijing have unveiled a new strategy to fix this issue. The team adapted a well-known computer-vision algorithm used for sharpening photos and, for the first time, applied it to astronomical images from ground-based telescopes. The researchers also trained the artificial intelligence (AI) algorithm on data simulated to match the Vera C. Rubin Observatory’s imaging parameters, so, when the observatory opens next year, the tool will be instantly compatible.
While astrophysicists already use technologies to remove blur, the adapted AI-driven algorithm works faster and produces more realistic images than current technologies. The resulting images are blur-free and truer to life. They also are beautiful — although that’s not the technology’s purpose.
“Photography’s goal is often to get a pretty, nice-looking image,” said Northwestern’s Emma Alexander, the study’s senior author. “But astronomical images are used for science. By cleaning up images in the right way, we can get more accurate data. The algorithm removes the atmosphere computationally, enabling physicists to obtain better scientific measurements. At the end of the day, the images do look better as well.”
The research will be published March 30 in the Monthly Notices of the Royal Astronomical Society.
Alexander is an assistant professor of computer science at Northwestern’s McCormick School of Engineering, where she runs the Bio Inspired Vision Lab. She co-led the new study with Tianao Li, an undergraduate in electrical engineering at Tsinghua University and a research intern in Alexander’s lab.

When light emanates from distant stars, planets and galaxies, it travels through Earth’s atmosphere before it hits our eyes. Not only does our atmosphere block out certain wavelengths of light, it also distorts the light that reaches Earth. Even clear night skies still contain moving air that affects light passing through it. That’s why stars twinkle and why the best ground-based telescopes are located at high altitudes where the atmosphere is thinnest.
“It’s a bit like looking up from the bottom of a swimming pool,” Alexander said. “The water pushes light around and distorts it. The atmosphere is, of course, much less dense, but it’s a similar concept.”
The blur becomes an issue when astrophysicists analyze images to extract cosmological data. By studying the apparent shapes of galaxies, scientists can detect the gravitational effects of large-scale cosmological structures, which bend light on its way to our planet. This can cause an elliptical galaxy to appear rounder or more stretched than it really is. But atmospheric blur smears the image in a way that warps the galaxy shape. Removing the blur enables scientists to collect accurate shape data.
“Slight differences in shape can tell us about gravity in the universe,” Alexander said. “These differences are already difficult to detect. If you look at an image from a ground-based telescope, a shape might be warped. It’s hard to know if that’s because of a gravitational effect or the atmosphere.”
To tackle this challenge, Alexander and Li combined an optimization algorithm with a deep-learning network trained on astronomical images. Among the training images, the team included simulated data that matches the Rubin Observatory’s expected imaging parameters. The resulting tool produced images with 38.6% less error compared to classic methods for removing blur and 7.4% less error compared to modern methods.

When the Rubin Observatory officially opens next year, its telescopes will begin a decade-long deep survey across an enormous portion of the night sky. Because the researchers trained the new tool on data specifically designed to simulate Rubin’s upcoming images, it will be able to help analyze the survey’s highly anticipated data.
For astronomers interested in using the tool, the open-source, user-friendly code and accompanying tutorials are available online.
“Now we pass off this tool, putting it into the hands of astronomy experts,” Alexander said. “We think this could be a valuable resource for sky surveys to obtain the most realistic data possible.”
The study, “Galaxy image deconvolution for weak gravitational lensing with unrolled plug-and-play ADMM,” used computational resources from the Computational Photography Lab at Northwestern University. More

A new artificial intelligence tool can predict the functions of enzymes based on their amino acid sequences, even when the enzymes are unstudied or poorly understood. The researchers said the AI tool, dubbed CLEAN, outperforms the leading state-of-the-art tools in accuracy, reliability and sensitivity. Better understanding of enzymes and their functions would be a boon for research in genomics, chemistry, industrial materials, medicine, pharmaceuticals and more.
“Just like ChatGPT uses data from written language to create predictive text, we are leveraging the language of proteins to predict their activity,” said study leader Huimin Zhao, a University of Illinois Urbana-Champaign professor of chemical and biomolecular engineering. “Almost every researcher, when working with a new protein sequence, wants to know right away what the protein does. In addition, when making chemicals for any application — biology, medicine, industry — this tool will help researchers quickly identify the proper enzymes needed for the synthesis of chemicals and materials.”
The researchers will publish their findings in the journal Science and make CLEAN accessible online March 31.
With advances in genomics, many enzymes have been identified and sequenced, but scientists have little or no information about what those enzymes do, said Zhao, a member of the Carl R. Woese Institute for Genomic Biology at Illinois.
Other computational tools try to predict enzyme functions. Typically, they attempt to assign an enzyme commission number — an ID code that indicates what kind of reaction an enzyme catalyzes — by comparing a queried sequence with a catalog of known enzymes and finding similar sequences. However, these tools don’t work as well with less-studied or uncharacterized enzymes, or with enzymes that perform multiple jobs, Zhao said.
“We are not the first one to use AI tools to predict enzyme commission numbers, but we are the first one to use this new deep-learning algorithm called contrastive learning to predict enzyme function. We find that this algorithm works much better than the AI tools that are used by others,” Zhao said. “We cannot guarantee everyone’s product will be correctly predicted, but we can get higher accuracy than the other two or other three methods.”
The researchers verified their tool experimentally with both computational and in vitro experiments. They found that not only could the tool predict the function of previously uncharacterized enzymes, it also corrected enzymes mislabeled by the leading software and correctly identified enzymes with two or more functions.

Zhao’s group is making CLEAN accessible online for other researchers seeking to characterize an enzyme or determine whether an enzyme could catalyze a desired reaction.
“We hope that this tool will be used widely by the broad research community,” Zhao said. “With the web interface, researchers can just enter the sequence in a search box, like a search engine, and see the results.”
Zhao said the group plans to expand the AI behind CLEAN to characterize other proteins, such as binding proteins. The team also hopes to further develop the machine-learning algorithms so that a user could search for a desired reaction and the AI would point to a proper enzyme for the job.
“There are a lot of uncharacterized binding proteins, such as receptors and transcription factors. We also want to predict their functions as well,” Zhao said. “We want to predict the functions of all proteins so that we can know all the proteins a cell has and better study or engineer the whole cell for biotechnology or biomedical applications.”
The National Science Foundation supported this work through the Molecule Maker Lab Institute, an AI Research Institute Zhao leads.
Further information: https://moleculemaker.org/alphasynthesis/ More

According to the U.S. Centers for Disease Control, one out of every three adults in the United States has prediabetes, a condition marked by elevated blood sugar levels that could lead to the development of Type 2 diabetes. The good news is that, if detected early, prediabetes can be reversed through lifestyle changes such as improved diet and exercise. The bad news? Eight out of 10 Americans with prediabetes don’t know that they have it, putting them at increased risk of developing diabetes as well as disease complications that include heart disease, kidney failure and vision loss.
Current screening methods typically involve a visit to a health care facility for laboratory testing and/or the use of a portable glucometer for at-home testing, meaning access and cost may be barriers to more widespread screening. But researchers at the University of Washington may have found the sweet spot when it comes to increasing early detection of prediabetes. The team developed GlucoScreen, a new system that leverages the capacitive touch sensing capabilities of any smartphone to measure blood glucose levels without the need for a separate reader.
The researchers describe GlucoScreen in a new paper published March 28 in the Proceedings of the Association for Computing Machinery on Interactive, Mobile, Wearable and Ubiquitous Technologies.
The researchers’ results suggest GlucoScreen’s accuracy is comparable to that of standard glucometer testing. The team found the system to be accurate at the crucial threshold between a normal blood glucose level, at or below 99 mg/dL, and prediabetes, defined as a blood glucose level between 100 and 125 mg/dL. This approach could make glucose testing less costly and more accessible — particularly for one-time screening of a large population.
“In conventional screening a person applies a drop of blood to a test strip, where the blood reacts chemically with the enzymes on the strip. A glucometer is used to analyze that reaction and deliver a blood glucose reading,” said lead author Anandghan Waghmare, a UW doctoral student in the Paul G. Allen School of Computer Science & Engineering. “We took the same test strip and added inexpensive circuitry that communicates data generated by that reaction to any smartphone through simulated tapping on the screen. GlucoScreen then processes the data and displays the result right on the phone, alerting the person if they are at risk so they know to follow up with their physician.”
Specifically, the GlucoScreen test strip samples the amplitude of the electrochemical reaction that occurs when a blood sample mixes with enzymes five times each second.

The strip then transmits the amplitude data to the phone through a series of touches at variable speeds using a technique called “pulse-width modulation.” The term “pulse width” refers to the distance between peaks in the signal — in this case, the length between taps. Each pulse width represents a value along the curve. The greater the distance between taps for a particular value, the higher the amplitude associated with the electrochemical reaction on the strip.
“You communicate with your phone by tapping the screen with your finger,” Waghmare said. “That’s basically what the strip is doing, only instead of a single tap to produce a single action, it’s doing multiple taps at varying speeds. It’s comparable to how Morse code transmits information through tapping patterns.”
The advantage of this technique is that it does not require complicated electronic components. This minimizes the cost to manufacture the strip and the power required for it to operate compared to more conventional communication methods, like Bluetooth and WiFi. All data processing and computation occurs on the phone, which simplifies the strip and further reduces the cost.
The test strip also doesn’t need batteries. It uses photodiodes instead to draw what little power it needs from the phone’s flash.
The flash is automatically engaged by the GlucoScreen app, which walks the user through each step of the testing process. First, a user affixes each end of the test strip to the front and back of the phone as directed. Next, they prick their finger with a lancet, as they would in a conventional test, and apply a drop of blood to the biosensor attached to the test strip. After the data is transmitted from the strip to the phone, the app applies machine learning to analyze the data and calculate a blood glucose reading.

That stage of the process is similar to that performed on a commercial glucometer. What sets GlucoScreen apart, in addition to its novel touch technique, is its universality.
“Because we use the built-in capacitive touch screen that’s present in every smartphone, our solution can be easily adapted for widespread use. Additionally, our approach does not require low-level access to the capacitive touch data, so you don’t have to access the operating system to make GlucoScreen work,” said co-author Jason Hoffman, a UW doctoral student in the Allen School. “We’ve designed it to be ‘plug and play.’ You don’t need to root the phone — in fact, you don’t need to do anything with the phone, other than install the app. Whatever model you have, it will work off the shelf.”
The researchers evaluated their approach using a combination of in vitro and clinical testing. Due to the COVID-19 pandemic, they had to delay the latter until 2021 when, on a trip home to India, Waghmare connected with Dr. Shailesh Pitale at Dew Medicare and Trinity Hospital. Upon learning about the UW project, Dr. Pitale agreed to facilitate a clinical study involving 75 consenting patients who were already scheduled to have blood drawn for a laboratory blood glucose test. Using that laboratory test as the ground truth, Waghmare and the team evaluated GlucoScreen’s performance against that of a conventional strip and glucometer.
Given how common prediabetes and diabetes are globally, this type of technology has the potential to change clinical care, the researchers said.
“One of the barriers I see in my clinical practice is that many patients can’t afford to test themselves, as glucometers and their test strips are too expensive. And, it’s usually the people who most need their glucose tested who face the biggest barriers,” said co-author Dr. Matthew Thompson, UW professor of both family medicine in the UW School of Medicine and global health. “Given how many of my patients use smartphones now, a system like GlucoScreen could really transform our ability to screen and monitor people with prediabetes and even diabetes.”
GlucoScreen is presently a research prototype. Additional user-focused and clinical studies, along with alterations to how test strips are manufactured and packaged, would be required before the system could be made widely available, the team said.
But, the researchers added, the project demonstrates how we have only begun to tap into the potential of smartphones as a health screening tool.
“Now that we’ve shown we can build electrochemical assays that can work with a smartphone instead of a dedicated reader, you can imagine extending this approach to expand screening for other conditions,” said senior author Shwetak Patel, the Washington Research Foundation Entrepreneurship Endowed Professor in Computer Science & Engineering and Electrical & Computer Engineering at the UW.
Additional co-authors are Farshid Salemi Parizi, a former UW doctoral student in electrical and computer engineering who is now a senior machine learning engineer at OctoML, and Yuntao Wang, a research professor at Tsinghua University and former visiting professor at the Allen School. This research was funded in part by the Bill & Melinda Gates Foundation. More

A new study led by Worcester Polytechnic Institute (WPI) brings into sharper focus the structural details of the COVID-19 virus, revealing an elliptical shape that “breathes,” or changes shape, as it moves in the body. The discovery, which could lead to new antiviral therapies for the disease and quicker development of vaccines, is featured in the April edition of the peer-reviewed Cell Press structural biology journal Structure.
“This is critical knowledge we need to fight future pandemics,” said Dmitry Korkin, Harold L. Jurist ’61 and Heather E. Jurist Dean’s Professor of Computer Science and lead researcher on the project. “Understanding the SARS-COV-2 virus envelope should allow us to model the actual process of the virus attaching to the cell and apply this knowledge to our understanding of the therapies at the molecular level. For instance, how can the viral activity be inhibited by antiviral drugs? How much antiviral blocking is needed to prevent virus-to-host interaction? We don’t know. But this is the best thing we can do right now — to be able to simulate actual processes.”
Feeding genetic sequencing information and massive amounts of real-world data about the pandemic virus into a supercomputer in Texas, Korkin and his team, working in partnership with a group led by Siewert-Jan Marrink at the University of Groningen, Netherlands, produced a computational model of the virus’s envelope, or outer shell, in “near atomistic detail” that had until now been beyond the reach of even the most powerful microscopes and imaging techniques.
Essentially, the computer used structural bioinformatics and computational biophysics to create its own picture of what the SARS-COV-2 particle looks like. And that picture showed that the virus is more elliptical than spherical and can change its shape. Korkin said the work also led to a better understanding of the M proteins in particular: underappreciated and overlooked components of the virus’s envelope.
The M proteins form entities called dimers with a copy of each other, and play a role in the particle’s shape-shifting by keeping the structure flexible overall while providing a triangular mesh-like structure on the interior that makes it remarkably resilient, Korkin said. In contrast, on the exterior, the proteins assemble into mysterious filament-like structures that have puzzled scientists who have seen Korkin’s results, and will require further study.
Korkin said the structural model developed by the researchers expands what was already known about the envelope architecture of the SARS-COV-2 virus and previous SARS- and MERS-related outbreaks. The computational protocol used to create the model could also be applied to more rapidly model future coronaviruses, he said. A clearer picture of the virus’ structure could reveal crucial vulnerabilities.
“The envelope properties of SARS-COV-2 are likely to be similar to other coronaviruses,” he said. “Eventually, knowledge about the properties of coronavirus membrane proteins could lead to new therapies and vaccines for future viruses.”
The new findings published in Structure were three years in the making and built upon Korkin’s work in the early days of the pandemic to provide the first 3D roadmap of the virus, based on genetic sequence information from the first isolated strain in China. More

When multiple drones are working together in the same airspace, perhaps spraying pesticide over a field of corn, there’s a risk they might crash into each other.
To help avoid these costly crashes, MIT researchers presented a system called MADER in 2020. This multiagent trajectory-planner enables a group of drones to formulate optimal, collision-free trajectories. Each agent broadcasts its trajectory so fellow drones know where it is planning to go. Agents then consider each other’s trajectories when optimizing their own to ensure they don’t collide.
But when the team tested the system on real drones, they found that if a drone doesn’t have up-to-date information on the trajectories of its partners, it might inadvertently select a path that results in a collision. The researchers revamped their system and are now rolling out Robust MADER, a multiagent trajectory planner that generates collision-free trajectories even when communications between agents are delayed.
“MADER worked great in simulations, but it hadn’t been tested in hardware. So, we built a bunch of drones and started flying them. The drones need to talk to each other to share trajectories, but once you start flying, you realize pretty quickly that there are always communication delays that introduce some failures,” says Kota Kondo, an aeronautics and astronautics graduate student.
The algorithm incorporates a delay-check step during which a drone waits a specific amount of time before it commits to a new, optimized trajectory. If it receives additional trajectory information from fellow drones during the delay period, it might abandon its new trajectory and start the optimization process over again.
When Kondo and his collaborators tested Robust MADER, both in simulations and flight experiments with real drones, it achieved a 100 percent success rate at generating collision-free trajectories. While the drones’ travel time was a bit slower than it would be with some other approaches, no other baselines could guarantee safety.

“If you want to fly safer, you have to be careful, so it is reasonable that if you don’t want to collide with an obstacle, it will take you more time to get to your destination. If you collide with something, no matter how fast you go, it doesn’t really matter because you won’t reach your destination,” Kondo says.
Kondo wrote the paper with Jesus Tordesillas, a postdoc; Parker C. Lusk, a graduate student; Reinaldo Figueroa, Juan Rached, and Joseph Merkel, MIT undergraduates; and senior author Jonathan P. How, the Richard C. Maclaurin Professor of Aeronautics and Astronautics and a member of the MIT-IBM Watson AI Lab. The research will be presented at the International Conference on Robots and Automation.
Planning trajectories
MADER is an asynchronous, decentralized, multiagent trajectory-planner. This means that each drone formulates its own trajectory and that, while all agents must agree on each new trajectory, they don’t need to agree at the same time. This makes MADER more scalable than other approaches, since it would be very difficult for thousands of drones to agree on a trajectory simultaneously. Due to its decentralized nature, the system would also work better in real-world environments where drones may fly far from a central computer.
With MADER, each drone optimizes a new trajectory using an algorithm that incorporates the trajectories it has received from other agents. By continually optimizing and broadcasting their new trajectories, the drones avoid collisions.

But perhaps one agent shared its new trajectory several seconds ago, but a fellow agent didn’t receive it right away because the communication was delayed. In real-world environments, signals are often delayed by interference from other devices or environmental factors like stormy weather. Due to this unavoidable delay, a drone might inadvertently commit to a new trajectory that sets it on a collision course.
Robust MADER prevents such collisions because each agent has two trajectories available. It keeps one trajectory that it knows is safe, which it has already checked for potential collisions. While following that original trajectory, the drone optimizes a new trajectory but does not commit to the new trajectory until it completes a delay-check step.
During the delay-check period, the drone spends a fixed amount of time repeatedly checking for communications from other agents to see if its new trajectory is safe. If it detects a potential collision, it abandons the new trajectory and starts the optimization process over again.
The length of the delay-check period depends on the distance between agents and environmental factors that could hamper communications, Kondo says. If the agents are many miles apart, for instance, then the delay-check period would need to be longer.
Completely collision-free
The researchers tested their new approach by running hundreds of simulations in which they artificially introduced communication delays. In each simulation, Robust MADER was 100 percent successful at generating collision-free trajectories, while all the baselines caused crashes.
The researchers also built six drones and two aerial obstacles and tested Robust MADER in a multiagent flight environment. They found that, while using the original version of MADER in this environment would have resulted in seven collisions, Robust MADER did not cause a single crash in any of the hardware experiments.
“Until you actually fly the hardware, you don’t know what might cause a problem. Because we know that there is a difference between simulations and hardware, we made the algorithm robust, so it worked in the actual drones, and seeing that in practice was very rewarding,” Kondo says.
Drones were able to fly 3.4 meters per second with Robust MADER, although they had a slightly longer average travel time than some baselines. But no other method was perfectly collision-free in every experiment.
In the future, Kondo and his collaborators want to put Robust MADER to the test outdoors, where many obstacles and types of noise can affect communications. They also want to outfit drones with visual sensors so they can detect other agents or obstacles, predict their movements, and include that information in trajectory optimizations.
This work was supported by Boeing Research and Technology. More

Artificial intelligence technologies like ChatGPT are seemingly doing everything these days: writing code, composing music, and even creating images so realistic you’ll think they were taken by professional photographers. Add thinking and responding like a human to the conga line of capabilities. A recent study from BYU proves that artificial intelligence can respond to complex survey questions just like a real human.
To determine the possibility of using artificial intelligence as a substitute for human responders in survey-style research, a team of political science and computer science professors and graduate students at BYU tested the accuracy of programmed algorithms of a GPT-3 language model — a model that mimics the complicated relationship between human ideas, attitudes, and sociocultural contexts of subpopulations.
In one experiment, the researchers created artificial personas by assigning the AI certain characteristics like race, age, ideology, and religiosity; and then tested to see if the artificial personas would vote the same as humans did in 2012, 2016, and 2020 U.S. presidential elections. Using the American National Election Studies (ANES) for their comparative human database, they found a high correspondence between how the AI and humans voted.
“I was absolutely surprised to see how accurately it matched up,” said David Wingate, BYU computer science professor, and co-author on the study. “It’s especially interesting because the model wasn’t trained to do political science — it was just trained on a hundred billion words of text downloaded from the internet. But the consistent information we got back was so connected to how people really voted.”
In another experiment, they conditioned artificial personas to offer responses from a list of options in an interview-style survey, again using the ANES as their human sample. They found high similarity between nuanced patterns in human and AI responses.
This innovation holds exciting prospects for researchers, marketers, and pollsters. Researchers envision a future where artificial intelligence is used to craft better survey questions, refining them to be more accessible and representative; and even simulate populations that are difficult to reach. It can be used to test surveys, slogans, and taglines as a precursor to focus groups.
“We’re learning that AI can help us understand people better,” said BYU political science professor Ethan Busby. “It’s not replacing humans, but it is helping us more effectively study people. It’s about augmenting our ability rather than replacing it. It can help us be more efficient in our work with people by allowing us to pre-test our surveys and our messaging.”
And while the expansive possibilities of large language models are intriguing, the rise of artificial intelligence poses a host of questions — how much does AI really know? Which populations will benefit from this technology and which will be negatively impacted? And how can we protect ourselves from scammers and fraudsters who will manipulate AI to create more sophisticated phishing scams?
While much of that is still to be determined, the study lays out a set of criteria that future researchers can use to determine how accurate an AI model is for different subject areas.
“We’re going to see positive benefits because it’s going to unlock new capabilities,” said Wingate, noting that AI can help people in many different jobs be more efficient. “We’re also going to see negative things happen because sometimes computer models are inaccurate and sometimes they’re biased. It will continue to churn society.”
Busby says surveying artificial personas shouldn’t replace the need to survey real people and that academics and other experts need to come together to define the ethical boundaries of artificial intelligence surveying in research related to social science. More

Everyone is talking about the newest AI and the power of neural networks, forgetting that software is limited by the hardware on which it runs. But it is hardware, says USC Professor of Electrical and Computer Engineering Joshua Yang, that has become “the bottleneck.” Now, Yang’s new research with collaborators might change that. They believe that they have developed a new type of chip with the best memory of any chip thus far for edge AI (AI in portable devices).
For approximately the past 30 years, while the size of the neural networks needed for AI and data science applications doubled every 3.5 months, the hardware capability needed to process them doubled only every 3.5 years. According to Yang, hardware presents a more and more severe problem for which few have patience.
Governments, industry, and academia are trying to address this hardware challenge worldwide. Some continue to work on hardware solutions with silicon chips, while others are experimenting with new types of materials and devices. Yang’s work falls into the middle — focusing on exploiting and combining the advantages of the new materials and traditional silicon technology that could support heavy AI and data science computation.
Their new paper in Nature focuses on the understanding of fundamental physics that leads to a drastic increase in memory capacity needed for AI hardware. The team led by Yang, with researchers from USC (including Han Wang’s group), MIT, and the University of Massachusetts, developed a protocol for devices to reduce “noise” and demonstrated the practicality of using this protocol in integrated chips. This demonstration was made at TetraMem, a startup company co-founded by Yang and his co-authors (Miao Hu, Qiangfei Xia, and Glenn Ge), to commercialize AI acceleration technology. According to Yang, this new memory chip has the highest information density per device (11 bits) among all types of known memory technologies thus far. Such small but powerful devices could play a critical role in bringing incredible power to the devices in our pockets. The chips are not just for memory but also for the processor. And millions of them in a small chip, working in parallel to rapidly run your AI tasks, could only require a small battery to power it.
The chips that Yang and his colleagues are creating combine silicon with metal oxide memristors in order to create powerful but low-energy intensive chips. The technique focuses on using the positions of atoms to represent information rather than the number of electrons (which is the current technique involved in computations on chips). The positions of the atoms offer a compact and stable way to store more information in an analog, instead of digital fashion. Moreover, the information can be processed where it is stored instead of being sent to one of the few dedicated ‘processors,’ eliminating the so-called ‘von Neumann bottleneck’ existing in current computing systems. In this way, says Yang, computing for AI is “more energy efficient with a higher throughput.”
How it works
Yang explains that electrons which are manipulated in traditional chips, are “light.” And this lightness, makes them prone to moving around and being more volatile. Instead of storing memory through electrons, Yang and collaborators are storing memory in full atoms. Here is why this memory matters. Normally, says Yang, when one turns off a computer, the information memory is gone — but if you need that memory to run a new computation and your computer needs the information all over again, you have lost both time and energy. This new method, focusing on activating atoms rather than electrons, does not require battery power to maintain stored information. Similar scenarios happen in AI computations, where a stable memory capable of high information density is crucial. Yang imagines this new tech that may enable powerful AI capability in edge devices, such as Google Glasses, which he says previously suffered from a frequent recharging issue.
Further, by converting chips to rely on atoms as opposed to electrons, chips become smaller. Yang adds that with this new method, there is more computing capacity at a smaller scale. And this method, he says, could offer “many more levels of memory to help increase information density.”
To put it in context, right now, ChatGPT is running on a cloud. The new innovation, followed by some further development, could put the power of a mini version of ChatGPT in everyone’s personal device. It could make such high-powered tech more affordable and accessible for all sorts of applications. More