More stories

  • in

    Machine learning guarantees robots' performance in unknown territory

    A small drone takes a test flight through a space filled with randomly placed cardboard cylinders acting as stand-ins for trees, people or structures. The algorithm controlling the drone has been trained on a thousand simulated obstacle-laden courses, but it’s never seen one like this. Still, nine times out of 10, the pint-sized plane dodges all the obstacles in its path.
    This experiment is a proving ground for a pivotal challenge in modern robotics: the ability to guarantee the safety and success of automated robots operating in novel environments. As engineers increasingly turn to machine learning methods to develop adaptable robots, new work by Princeton University researchers makes progress on such guarantees for robots in contexts with diverse types of obstacles and constraints.
    “Over the last decade or so, there’s been a tremendous amount of excitement and progress around machine learning in the context of robotics, primarily because it allows you to handle rich sensory inputs,” like those from a robot’s camera, and map these complex inputs to actions, said Anirudha Majumdar, an assistant professor of mechanical and aerospace engineering at Princeton.
    However, robot control algorithms based on machine learning run the risk of overfitting to their training data, which can make algorithms less effective when they encounter inputs that differ from those they were trained on. Majumdar’s Intelligent Robot Motion Lab addressed this challenge by expanding the suite of available tools for training robot control policies, and quantifying the likely success and safety of robots performing in novel environments.
    In three new papers, the researchers adapted machine learning frameworks from other arenas to the field of robot locomotion and manipulation. They turned to generalization theory, which is typically used in contexts that map a single input onto a single output, such as automated image tagging. The new methods are among the first to apply generalization theory to the more complex task of making guarantees on robots’ performance in unfamiliar settings. While other approaches have provided such guarantees under more restrictive assumptions, the team’s methods offer more broadly applicable guarantees on performance in novel environments, said Majumdar.
    In the first paper, a proof of principle for applying the machine learning frameworks, the team tested their approach in simulations that included a wheeled vehicle driving through a space filled with obstacles, and a robotic arm grasping objects on a table. They also validated the technique by assessing the obstacle avoidance of a small drone called a Parrot Swing (a combination quadcopter and fixed-wing airplane) as it flew down a 60-foot-long corridor dotted with cardboard cylinders. The guaranteed success rate of the drone’s control policy was 88.4%, and it avoided obstacles in 18 of 20 trials (90%).

    advertisement

    The work, published Oct. 3 in the International Journal of Robotics Research, was coauthored by Majumdar; Alec Farid, a graduate student in mechanical and aerospace engineering; and Anoopkumar Sonar, a computer science concentrator from Princeton’s Class of 2021.
    When applying machine learning techniques from other areas to robotics, said Farid, “there are a lot of special assumptions you need to satisfy, and one of them is saying how similar the environments you’re expecting to see are to the environments your policy was trained on. In addition to showing that we can do this in the robotic setting, we also focused on trying to expand the types of environments that we could provide a guarantee for.”
    “The kinds of guarantees we’re able to give range from about 80% to 95% success rates on new environments, depending on the specific task, but if you’re deploying [an unmanned aerial vehicle] in a real environment, then 95% probably isn’t good enough,” said Majumdar. “I see that as one of the biggest challenges, and one that we are actively working on.”
    Still, the team’s approaches represent much-needed progress on generalization guarantees for robots operating in unseen environments, said Hongkai Dai, a senior research scientist at the Toyota Research Institute in Los Altos, California.
    “These guarantees are paramount to many safety-critical applications, such as self-driving cars and autonomous drones, where the training set cannot cover every possible scenario,” said Dai, who was not involved in the research. “The guarantee tells us how likely it is that a policy can still perform reasonably well on unseen cases, and hence establishes confidence on the policy, where the stake of failure is too high.”
    In two other papers, to be presented Nov. 18 at the virtual Conference on Robot Learning, the researchers examined additional refinements to bring robot control policies closer to the guarantees that would be needed for real-world deployment. One paper used imitation learning, in which a human “expert” provides training data by manually guiding a simulated robot to pick up various objects or move through different spaces with obstacles. This approach can improve the success of machine learning-based control policies.

    advertisement

    To provide the training data, lead author Allen Ren, a graduate student in mechanical and aerospace engineering, used a 3D computer mouse to control a simulated robotic arm tasked with grasping and lifting drinking mugs of various sizes, shapes and materials. Other imitation learning experiments involved the arm pushing a box across a table, and a simulation of a wheeled robot navigating around furniture in a home-like environment.
    The researchers deployed the policies learned from the mug-grasping and box-pushing tasks on a robotic arm in the laboratory, which was able to pick up 25 different mugs by grasping their rims between its two finger-like grippers — not holding the handle as a human would. In the box-pushing example, the policy achieved 93% success on easier tasks and 80% on harder tasks.
    “We have a camera on top of the table that sees the environment and takes a picture five times per second,” said Ren. “Our policy training simulation takes this image and outputs what kind of action the robot should take, and then we have a controller that moves the arm to the desired locations based on the output of the model.”
    A third paper demonstrated the development of vision-based planners that provide guarantees for flying or walking robots to carry out planned sequences of movements through diverse environments. Generating control policies for planned movements brought a new problem of scale — a need to optimize vision-based policies with thousands, rather than hundreds, of dimensions.
    “That required coming up with some new algorithmic tools for being able to tackle that dimensionality and still be able to give strong generalization guarantees,” said lead author Sushant Veer, a postdoctoral research associate in mechanical and aerospace engineering.
    A key aspect of Veer’s strategy was the use of motion primitives, in which a policy directs a robot to go straight or turn, for example, rather than specifying a torque or velocity for each movement. Narrowing the space of possible actions makes the planning process more computationally tractable, said Majumdar.
    Veer and Majumdar evaluated the vision-based planners on simulations of a drone navigating around obstacles and a four-legged robot traversing rough terrain with slopes as high as 35 degrees — “a very challenging problem that a lot of people in robotics are still trying to solve,” said Veer.
    In the study, the legged robot achieved an 80% success rate on unseen test environments. The researchers are working to further improve their policies’ guarantees, as well as assessing the policies’ performance on real robots in the laboratory.
    The work was supported in part by the U.S. Office of Naval Research, the National Science Foundation, a Google Faculty Research Award and an Amazon Research Award. More

  • in

    AI tool may predict movies' future ratings

    Movie ratings can determine a movie’s appeal to consumers and the size of its potential audience. Thus, they have an impact on a film’s bottom line. Typically, humans do the tedious task of manually rating a movie based on viewing the movie and making decisions on the presence of violence, drug abuse and sexual content.
    Now, researchers at the USC Viterbi School of Engineering, armed with artificial intelligence tools, can rate a movie’s content in a matter of seconds, based on the movie script and before a single scene is shot. Such an approach could allow movie executives the ability to design a movie rating in advance and as desired, by making the appropriate edits on a script and before the shooting of a single scene. Beyond the potential financial impact, such instantaneous feedback would allow storytellers and decision-makers to reflect on the content they are creating for the public and the impact such content might have on viewers.
    Using artificial intelligence applied to scripts, Shrikanth Narayanan, University Professor and Niki & C. L. Max Nikias Chair in Engineering, and a team of researchers from the Signal Analysis and Interpretation Lab (SAIL) at USC Viterbi, have demonstrated that linguistic cues can effectively signal behaviors on violent acts, drug abuse and sexual content (actions that are often the basis for a film’s ratings) about to be taken by a film’s characters.
    Method:
    Using 992 movie scripts that included violent, substance-abuse and sexual content, as determined by Common Sense Media, a non-profit organization that rates and makes recommendations for families and schools, the SAIL research team trained artificial intelligence to recognize corresponding risk behaviors, patterns and language.
    The AI tool created receives as input all the script, processes it through a neural network and scans it for semantics and sentiment expressed. In the process, it classifies sentences and phrases as positive, negative, aggressive and other descriptors. The AI tool automatically classifies words and phrases into three categories: violence, drug abuse and sexual content.

    advertisement

    Victor Martinez, a doctoral candidate in computer science at USC Viterbi and the lead researcher on the study, which will appear in The Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing said, “Our model looks at the movie script, rather than the actual scenes, including e.g. sounds like a gunshot or explosion that occur later in the production pipeline. This has the benefit of providing a rating long before production to help filmmakers decide e.g. the degree of violence and whether it needs to be toned down.”
    The research team also includes Narayanan, a professor of electrical and computer engineering, computer science and linguistics, Krishna Somandepalli, a Ph.D. candidate in Electrical and Computing Engineering at USC Viterbi, and Professor Yalda T. Uhls of UCLA’s Department of Psychology. They discovered many interesting connections between the portrayals of risky behaviors.
    “There seems to be a correlation in the amount of content in a typical film focused on substance abuse and the amount of sexual content. Whether intentionally or not, filmmakers seem to match the level of substance abuse-related content with sexually explicit content,” said Martinez.
    Another interesting pattern also emerged. “We found that filmmakers compensate for low levels of violence with joint portrayals of substance abuse and sexual content,” Martinez said.
    Moreover, while many movies contain depictions of rampant drug-abuse and sexual content, the researchers found it highly unlikely for a film to have high levels of all three risky behaviors, perhaps because of Motion Picture Association (MPA) standards.

    advertisement

    They also found an interesting connection between risk behaviors and MPA ratings. As sexual content increases, the MPA appears to put less emphasis on violence/substance-abuse content. Thus, regardless of violent and substance abuse content, a movie with a lot of sexual content will likely receive an R rating.
    Narayanan whose SAIL lab has pioneered the field of media informatics and applied natural language processing in order to bring awareness in the creative community about the nuances of storytelling, calls media “a rich avenue for studying human communication, interaction and behavior, since it provides a window into society.”
    “At SAIL, we are designing technologies and tools, based on AI, for all stakeholders in this creative business — the writers, film-makers and producers — to raise awareness about the varied important details associated in telling their story on film,” Narayanan said.
    “Not only are we interested in the perspective of the storytellers of the narratives they weave,” Narayanan said, “but also in understanding the impact on the audience and the ‘take-away’ from the whole experience. Tools like these will help raise societally-meaningful awareness, for example, through identifying negative stereotypes.”
    Added Martinez: “In the future, I’m interested in studying minorities and how they are represented, particularly in cases of violence, sex and drugs.” More

  • in

    Sensor experts invent supercool mini thermometer

    Researchers at the National Institute of Standards and Technology (NIST) have invented a miniature thermometer with big potential applications such as monitoring the temperature of processor chips in superconductor-based quantum computers, which must stay cold to work properly.
    NIST’s superconducting thermometer measures temperatures below 1 Kelvin (minus 272.15 ?C or minus 457.87 ?F), down to 50 milliKelvin (mK) and potentially 5 mK. It is smaller, faster and more convenient than conventional cryogenic thermometers for chip-scale devices and could be mass produced. NIST researchers describe the design and operation in a new journal paper.
    Just 2.5 by 1.15 millimeters in size, the new thermometer can be embedded in or stuck to another cryogenic microwave device to measure its temperature when mounted on a chip. The researchers used the thermometer to demonstrate fast, accurate measurements of the heating of a superconducting microwave amplifier.
    The technology is a spinoff of NIST’s custom superconducting sensors for telescope cameras, specifically microwave detectors delivered for the BLAST balloon.
    “This was a fun idea that quickly grew into something very helpful,” group leader Joel Ullom said. “The thermometer allows researchers to measure the temperature of a wide range of components in their test packages at very little cost and without introducing a large number of additional electrical connections. This has the potential to benefit researchers working in quantum computing or using low-temperature sensors in a wide range of fields.”
    The thermometer consists of a superconducting niobium resonator coated with silicon dioxide. The coating interacts with the resonator to shift the frequency at which it naturally vibrates. Scientists suspect this is due to atoms “tunneling” between two sites, a quantum-mechanical effect.
    The NIST thermometer is based on a new application of the principle that the natural frequency of the resonator depends on the temperature. The thermometer maps changes in frequency, as measured by electronics, to a temperature. By contrast, conventional thermometers for sub-Kelvin temperatures are based on electrical resistance. They require wiring routed to room-temperature electronics, adding complexity and potentially causing heating and interference.
    The NIST thermometer measures temperature in about 5 milliseconds (thousandths of a second), much faster than most conventional resistive thermometers at about one-tenth of a second. The NIST thermometers are also easy to fabricate in only a single process step. They can be mass produced, with more than 1,200 fitting on a 3-inch (approximately 75-millimeter) silicon wafer.

    Story Source:
    Materials provided by National Institute of Standards and Technology (NIST). Note: Content may be edited for style and length. More

  • in

    Time to rethink predicting pandemic infection rates?

    During the first months of the COVID-19 pandemic, Joseph Lee McCauley, a physics professor at the University of Houston, was watching the daily data for six countries and wondered if infections were really growing exponentially. By extracting the doubling times from the data, he became convinced they were.
    Doubling times and exponential growth go hand in hand, so it became clear to him that modeling based on past infections is impossible, because the rate changes unforeseeably from day to day due to social distancing and lockdown efforts. And the rate changes differ for each country based on the extent of their social distancing.
    In AIP Advances, from AIP Publishing, McCauley explains how he combined math in the form of Tchebychev’s inequality with a statistical ensemble to understand how macroscopic exponential growth with different daily rates arise from person-to-person disease infection.
    “Discretized ordinary chemical kinetic equations applied to infected, uninfected, and recovered parts of the population allowed me to organize the data, so I could separate the effects of social distancing and recoveries within daily infection rates,” McCauley said.
    Plateauing without peaking occurs if the recovery rate is too low, and the U.S., U.K., and Sweden fall into that category. Equations cannot be iterated to look into the future, because tomorrow’s rate is unknown until it unfolds.
    “Modelers tend to misapply the chemical kinetic equations as SIR (Susceptible, Infectious, or Recovered) or SEIR (Susceptible, Exposed, Infectious, or Recovered) models, because they are trying to generate future rates from past rates,” McCauley said. “But the past doesn’t allow you to use equations to predict the future in a pandemic, because social distancing changes the rates daily.”
    McCauley discovered he could make a forecast within five seconds via hand calculator that is as good as any computer model by simply using infection rates for today and yesterday.
    “Lockdowns and social distancing work,” said McCauley. “Compare Austria, Germany, Taiwan, Denmark, Finland, and several other countries that peaked in early April, with the U.S., U.K., Sweden, and others with no lockdown or half-hearted lockdowns — they’ve never even plateaued, much less peaked.”
    He stresses that forecasting cannot foresee peaking or even plateauing. Plateauing does not imply peaking, and if peaking occurs, there is nothing in the data to show when it will happen. It happens when the recovery rate is greater than the rate of new infections.
    “Social distancing and lockdowns reduce the infection rate but can’t cause peaking,” McCauley said. “Social distancing and recoveries are two separate terms within the daily kinetic rate equations.”
    The implication of this work is that research money could be better spent than on expensive epidemic modeling.
    “Politicians should know enough arithmetic to be given instruction on the implications,” McCauley said. “The effect of lockdowns and social distancing show up in the observed doubling times, and there is also a predicted doubling time based on two days, which serves as a good forecast of the future.” More

  • in

    In a pandemic, migration away from dense cities more effective than closing borders

    Pandemics are fueled, in part, by dense populations in large cities where networks of buildings, crowded sidewalks, and public transportation force people into tighter conditions. This contrasts with conditions in rural areas, where there is more space available per person.
    According to common sense, being in less crowded areas during a pandemic is safer. But small town mayors want to keep people safe, too, and migration of people from cities to rural towns brings concerns. During the COVID-19 pandemic, closing national borders and borders between states and regions has been prevalent. But does it really help?
    In a paper published in Chaos, by AIP Publishing, two researchers decided to put this hypothesis to the test and discover if confinement and travels bans are really effective ways to limit the spread of a pandemic disease. Specifically, they focused on the movement of people from larger cities to smaller ones and tested the results of this one-way migration.
    “Instead of taking mobility, or the lack of mobility, for granted, we decided to explore how an altered mobility would affect the spreading,” author Massimiliano Zanin said. “The real answer lies in the sign of the result. People always assume that closing borders is good. We found that it is almost always bad.”
    The model used by the authors is simplified, without many of the details that affect migration patterns and disease spread. But their focus on changes in population density indicates travel bans might be less effective than migration of people to less dense areas. The result was reduced spread of disease.
    Zanin and collaborator David Papo placed a hypothetical group of people in two locations and assumed their travel was in random movement patterns. They used SIR dynamics, which is common in epidemiological studies of disease movement. SIR stands for susceptible, infected, and recovered — classifications used to label groups in a simulation and track disease spread according to their interactions.
    They ran 10,000 iterations of the simulation to determine the resulting disease spread among people in two locations when migration is one way: from dense cities to less dense towns. They also studied the effect of “forced migration,” which moves healthy people out of dense cities at the onset of a pandemic.
    The results showed that while movement from big cities to small towns might be slightly less safe for the people in small towns, overall, for a global pandemic situation, this reduction in the density of highly populated areas is better for the majority of all people.
    “Collaboration between different governments and administrations is an essential ingredient towards controlling a pandemic, and one should consider the possibility of small-scale sacrifices to reach a global benefit,” Zanin said.

    Story Source:
    Materials provided by American Institute of Physics. Note: Content may be edited for style and length. More

  • in

    Quantum algorithm breakthrough

    Researchers led by City College of New York physicist Pouyan Ghaemi report the development of a quantum algorithm with the potential to study a class of many-electron quantums system using quantum computers. Their paper, entitled “Creating and Manipulating a Laughlin-Type ?=1/3 Fractional Quantum Hall State on a Quantum Computer with Linear Depth Circuits,” appears in the December issue of PRX Quantum, a journal of the American Physical Society.
    “Quantum physics is the fundamental theory of nature which leads to formation of molecules and the resulting matter around us,” said Ghaemi, assistant professor in CCNY’s Division of Science. “It is already known that when we have a macroscopic number of quantum particles, such as electrons in the metal, which interact with each other, novel phenomena such as superconductivity emerge.”
    However, until now, according to Ghaemi, tools to study systems with large numbers of interacting quantum particles and their novel properties have been extremely limited.
    “Our research has developed a quantum algorithm which can be used to study a class of many-electron quantum systems using quantum computers. Our algorithm opens a new venue to use the new quantum devices to study problems which are quite challenging to study using classical computers. Our results are new and motivate many follow up studies,” added Ghaemi.
    On possible applications for this advancement, Ghaemi, who’s also affiliated with the Graduate Center, CUNY noted: “Quantum computers have witnessed extensive developments during the last few years. Development of new quantum algorithms, regardless of their direct application, will contribute to realize applications of quantum computers.
    “I believe the direct application of our results is to provide tools to improve quantum computing devices. Their direct real-life application would emerge when quantum computers can be used for daily life applications.”
    His collaborators included scientists from: Western Washington University, University of California, Santa Barbara; Google AI Quantum and the University of Michigan, Ann Arbor.

    Story Source:
    Materials provided by City College of New York. Note: Content may be edited for style and length. More

  • in

    System brings deep learning to 'internet of things' devices

    Deep learning is everywhere. This branch of artificial intelligence curates your social media and serves your Google search results. Soon, deep learning could also check your vitals or set your thermostat. MIT researchers have developed a system that could bring deep learning neural networks to new — and much smaller — places, like the tiny computer chips in wearable medical devices, household appliances, and the 250 billion other objects that constitute the “internet of things” (IoT).
    The system, called MCUNet, designs compact neural networks that deliver unprecedented speed and accuracy for deep learning on IoT devices, despite limited memory and processing power. The technology could facilitate the expansion of the IoT universe while saving energy and improving data security.
    The research will be presented at next month’s Conference on Neural Information Processing Systems. The lead author is Ji Lin, a PhD student in Song Han’s lab in MIT’s Department of Electrical Engineering and Computer Science. Co-authors include Han and Yujun Lin of MIT, Wei-Ming Chen of MIT and National University Taiwan, and John Cohn and Chuang Gan of the MIT-IBM Watson AI Lab.
    The Internet of Things
    The IoT was born in the early 1980s. Grad students at Carnegie Mellon University, including Mike Kazar ’78, connected a Cola-Cola machine to the internet. The group’s motivation was simple: laziness. They wanted to use their computers to confirm the machine was stocked before trekking from their office to make a purchase. It was the world’s first internet-connected appliance. “This was pretty much treated as the punchline of a joke,” says Kazar, now a Microsoft engineer. “No one expected billions of devices on the internet.”
    Since that Coke machine, everyday objects have become increasingly networked into the growing IoT. That includes everything from wearable heart monitors to smart fridges that tell you when you’re low on milk. IoT devices often run on microcontrollers — simple computer chips with no operating system, minimal processing power, and less than one thousandth of the memory of a typical smartphone. So pattern-recognition tasks like deep learning are difficult to run locally on IoT devices. For complex analysis, IoT-collected data is often sent to the cloud, making it vulnerable to hacking.

    advertisement

    “How do we deploy neural nets directly on these tiny devices? It’s a new research area that’s getting very hot,” says Han. “Companies like Google and ARM are all working in this direction.” Han is too.
    With MCUNet, Han’s group codesigned two components needed for “tiny deep learning” — the operation of neural networks on microcontrollers. One component is TinyEngine, an inference engine that directs resource management, akin to an operating system. TinyEngine is optimized to run a particular neural network structure, which is selected by MCUNet’s other component: TinyNAS, a neural architecture search algorithm.
    System-algorithm codesign
    Designing a deep network for microcontrollers isn’t easy. Existing neural architecture search techniques start with a big pool of possible network structures based on a predefined template, then they gradually find the one with high accuracy and low cost. While the method works, it’s not the most efficient. “It can work pretty well for GPUs or smartphones,” says Lin. “But it’s been difficult to directly apply these techniques to tiny microcontrollers, because they are too small.”
    So Lin developed TinyNAS, a neural architecture search method that creates custom-sized networks. “We have a lot of microcontrollers that come with different power capacities and different memory sizes,” says Lin. “So we developed the algorithm [TinyNAS] to optimize the search space for different microcontrollers.” The customized nature of TinyNAS means it can generate compact neural networks with the best possible performance for a given microcontroller — with no unnecessary parameters. “Then we deliver the final, efficient model to the microcontroller,” say Lin.

    advertisement

    To run that tiny neural network, a microcontroller also needs a lean inference engine. A typical inference engine carries some dead weight — instructions for tasks it may rarely run. The extra code poses no problem for a laptop or smartphone, but it could easily overwhelm a microcontroller. “It doesn’t have off-chip memory, and it doesn’t have a disk,” says Han. “Everything put together is just one megabyte of flash, so we have to really carefully manage such a small resource.” Cue TinyEngine.
    The researchers developed their inference engine in conjunction with TinyNAS. TinyEngine generates the essential code necessary to run TinyNAS’ customized neural network. Any deadweight code is discarded, which cuts down on compile-time. “We keep only what we need,” says Han. “And since we designed the neural network, we know exactly what we need. That’s the advantage of system-algorithm codesign.” In the group’s tests of TinyEngine, the size of the compiled binary code was between 1.9 and five times smaller than comparable microcontroller inference engines from Google and ARM. TinyEngine also contains innovations that reduce runtime, including in-place depth-wise convolution, which cuts peak memory usage nearly in half. After codesigning TinyNAS and TinyEngine, Han’s team put MCUNet to the test.
    MCUNet’s first challenge was image classification. The researchers used the ImageNet database to train the system with labeled images, then to test its ability to classify novel ones. On a commercial microcontroller they tested, MCUNet successfully classified 70.7 percent of the novel images — the previous state-of-the-art neural network and inference engine combo was just 54 percent accurate. “Even a 1 percent improvement is considered significant,” says Lin. “So this is a giant leap for microcontroller settings.”
    The team found similar results in ImageNet tests of three other microcontrollers. And on both speed and accuracy, MCUNet beat the competition for audio and visual “wake-word” tasks, where a user initiates an interaction with a computer using vocal cues (think: “Hey, Siri”) or simply by entering a room. The experiments highlight MCUNet’s adaptability to numerous applications.
    “Huge potential”
    The promising test results give Han hope that it will become the new industry standard for microcontrollers. “It has huge potential,” he says.
    The advance “extends the frontier of deep neural network design even farther into the computational domain of small energy-efficient microcontrollers,” says Kurt Keutzer, a computer scientist at the University of California at Berkeley, who was not involved in the work. He adds that MCUNet could “bring intelligent computer-vision capabilities to even the simplest kitchen appliances, or enable more intelligent motion sensors.”
    MCUNet could also make IoT devices more secure. “A key advantage is preserving privacy,” says Han. “You don’t need to transmit the data to the cloud.”
    Analyzing data locally reduces the risk of personal information being stolen — including personal health data. Han envisions smart watches with MCUNet that don’t just sense users’ heartbeat, blood pressure, and oxygen levels, but also analyze and help them understand that information. MCUNet could also bring deep learning to IoT devices in vehicles and rural areas with limited internet access.
    Plus, MCUNet’s slim computing footprint translates into a slim carbon footprint. “Our big dream is for green AI,” says Han, adding that training a large neural network can burn carbon equivalent to the lifetime emissions of five cars. MCUNet on a microcontroller would require a small fraction of that energy. “Our end goal is to enable efficient, tiny AI with less computational resources, less human resources, and less data,” says Han. More