Bringing light into computers to accelerate AI and machine learning

Research, Electrical & Computer, Banner

BRINGING LIGHT INTO COMPUTERS TO ACCELERATE AI AND MACHINE LEARNINGIt might not be commonly known, but artificial intelligence and machine learning applications are commonplace today, performing a multitude of tasks for us behind the scenes. For example, AI and machine learning helps to interpret voice commands given to our phones and devices such as Alexa, recommends movies and music we might enjoy through services such as Netflix and Spotify, and is even driving autonomous vehicles. In the near future, the reach of AI and machine learning applications is expected to extend even further, to more complex tasks such as supporting space missionsand defense operations, and developing new drugs to treat disease.But the growing sophistication of AI and machine learning applications, as well as their implementation at such a large scale, demands a need for computing power which roughly doubles every three to four months. That’s much faster than Moore’s law (the observation that the number of transistors in a dense, integrated circuit doubles about every two years). Conventional computing paradigms and hardware platforms are having trouble keeping up. Also, cloud computing data centers used by AI and machine learning applications around the world currently gobble up an estimated 200-terawatt hours per year. That’s more than a small country. It’s easy to see that this energy consumption will come hand-in-hand with serious environmental consequences.To help address these challenges, UW ECE faculty members Sajjad Moazeni and Mo Li are leading a multi-institutional research team that recently received a four-year grant from the National Science Foundation to develop a new type of computer chip that uses laser light for AI and machine learning computation. This chip, called a “hybrid co-processing unit,” or HCU, stands to greatly accelerate the computing speed and efficiency of AI and machine learning applications, while at the same time reducing energy consumption. The computational power of the HCU will be over ten times greater than today’s most advanced silicon-based microprocessors of comparable size.“There is a need to shift the computing paradigm to something new,” said Moazeni, who is lead principal investigator of the project. “One of the most important and distinctive novelties in the work we are doing is that what we are proposing can very tightly get integrated with existing silicon-based microprocessors in today’s data centers. That is something very unique.”A new, scalable optical computing paradigmThe HCU combines traditional electronics with photonics, using light generated by lasers instead of electricity for data processing. The device does this by way of an optical computing core that includes phase-change material (a substance similar to what is in CD-ROMs and DVDs) to record information. This computing core can realize an optical neural network on the chip to accelerate computational speed in an ultracompact footprint, storing data on-chip using the phase-change material at essentially zero-power.“The HCU is a single-chip solution that can be integrated with today’s silicon-based microprocessors,” Moazeni said. “We call it ‘hybrid’ because we are co-optimizing the benefits of electronics, photonics and phase-change materials, all within one system.”The project builds on research by Moazeni, who is an expert in large-scale integrated photonics and microelectronics, as well as Li, who has been developing optical computing systems using phase-change materials at UW ECE. According to Moazeni and Li, this is the first time photonics and electronics have been so tightly integrated together in a single chip for the purpose of accelerating AI and machine learning computations.“Optical computing is best for data movement and linear computation, while traditional electronics are really good at digital computation and also implementing nonlinear algorithms, which optical computing cannot easily do,” Li said. “Our strategy combines the best of the two.”Other members of the research team are Nathan Youngblood, an assistant professor of electrical and computer engineering at the University of Pittsburgh, and Lei Jiang, an assistant professor of intelligent systems engineering at Indiana University Bloomington. Youngblood will work on designing electrically programmable, high density optical memory arrays for ultrafast optical computation, and Jiang will be focusing on optimizing the device for accelerating emerging AI and machine learning applications. What’s next?The research team is working toward combining the phase-change material with microelectronics circuitry at the Washington Nanofabrication Facility. This will be achieved through integrating the phase-change material with an advanced silicon photonic process fabricated at a commercial foundry. The method allows thousands of photonic elements and millions of transistors to be fabricated together in a cost-effective and scalable manner. The team will also be building computer models to simulate every aspect of the device.“We’ll start by modeling and putting together the full end-to-end model of the HCU, model the phase-change material, model the photonics and construct a new, unique framework on which we can simulate all of them together,” Moazeni said.By the end of the NSF grant in 2025, the research team expects to have a working, physical prototype. Then, the group will be poised to manufacture the device in larger quantities and at a scale capable of moving into the marketplace.What does that mean for the rest of us? Eventually, the work promises to translate into quicker response times and improved performance for any computer application that involves AI or machine learning (such as our phones, Alexa, Netflix and Spotify). It also will help make possible a significant reduction in energy consumption, making technology driven by AI and machine learning more environmentally friendly.“This is the first time that we’ll be bringing a non-traditional computing chip into the real world for practical applications, and I’m really excited about that,” Moazeni said. “It’s a realization of Moore’s law, which stated that eventually new materials would need to be brought into chip development in order to increase computing capacity and speed.”“Our technology will improve speed, performance and power consumption,” Li added. “And perhaps most importantly, it will help to put AI computing on a sustainable energy path.”By Wayne Gillam | UW ECE News

Snails carrying the world’s smallest computer help solve mass extinction survivor mystery

Research, Banner, Electrical & Computer

More than 50 species of tree snail in the South Pacific Society Islands were wiped out following the introduction of an alien predatory snail in the 1970s, but the white-shelled Partula hyalina survived.Now, thanks to a collaboration between University of Michigan biologists and engineers with the world’s smallest computer, scientists understand why: P. hyalina can tolerate more sunlight than its predator, so it was able to persist in sunlit forest edge habitats.“We were able to get data that nobody had been able to obtain,” said David Blaauw, the Kensall D. Wise Collegiate Professor of Electrical Engineering and Computer Science. “And that’s because we had a tiny computing system that was small enough to stick on a snail.”Most ecology and conservation studies involving data from sensors are done on vertebrate animals, which can carry larger and heavier devices than invertebrates. The current study not only offers insights into the conservation measures needed to ensure the survival of a species of snails, it points the way for future studies of very small animals through similar partnerships.“A lot of the coolest scientific work is done at the interface, where you have a classic problem and need to bring new approaches to find a solution,” said Diarmaid Ó Foighil, professor of Ecology and Evolutionary Biology (EEB) and Curator of the U-M Museum of Zoology.The Michigan Micro Mote (M3), considered the world’s smallest complete computer, was announced in 2014 by a team Blaauw co-led. This was its first field application.“The sensing computers are helping us understand how to protect endemic species on islands,” said Cindy Bick, who received a Ph.D. in ecology and evolutionary biology from U-M in 2018. “If we are able to map and protect these habitats through appropriate conservation measures, we can figure out ways to ensure the survival of the species.”P. hyalina is important culturally for Polynesians because of its unique color, making it attractive for use in shell leis and jewelry. Tree snails also play a vital role in island forest ecosystems, as the dominant group of native grazers.How Society Island snails were wiped outThe giant African land snail was introduced to the Society Islands, including Tahiti, to cultivate as a food source, but it became a major pest. To control its population, agricultural scientists introduced the rosy wolf snail in 1974. But unfortunately, most of the 61 known species of native Society Islands tree snails were easy prey for the rosy wolf snail. P. hyalina is one of only five survivors in the wild. Called the “Darwin finches of the snail world” for their island-bound diversity, the loss of so many Partula species is a blow to biologists studying evolution.“The endemic tree snails had never encountered a predator like the alien rosy wolf snail before it’s deliberate introduction. It can climb trees and very quickly drove most of the valley populations to local extinction,” said Ó FoighilIn 2015, Ó Foighil and Bick hypothesized that P. hyalina‘s distinctive white shell might give it an important advantage in forest edge habitats, by reflecting rather than absorbing light radiation levels that would be deadly to its darker-shelled predator. To test their idea, they needed to be able to track the light exposure levels P. hyalina and rosy wolf snails experienced in a typical day. Bick and Ó Foighil wanted to attach light sensors to the snails, but a system made using commercially available chips would have been too large for their purposes.Matching the technology to the application “In 2015, I was looking around to see who had the technology capable of tracking snails, and discovered that the people who can do this are also at Michigan. It was serendipitous,” said Bick. The system Bick found online was known as the Michigan Micro Mote (M3), which was created to go where other sensing devices could not. Measuring just 2x5x2mm, including packaging, it could be used to track animals the size of the rosy wolf snail (with an adult shell 3-7cm) in their natural habitat.But could it be altered to sense light? It was time to move the M3 from the laboratory to the real world, and adapt it for the special needs of their collaborators.“It was important to understand what the biologists were thinking and what they needed,” said Inhee Lee, an assistant professor of electrical and computer engineering at the University of Pittsburgh who received a Ph.D. from U-M electrical and computer engineering in 2014. Lee adapted the M3 for the study.The first step was to figure out how to measure the light intensity of the snails’ habitats. At the time, the team had just added an energy harvester to the M3 system to recharge the battery using tiny solar cells developed by Jamie Phillips.  Lee realized he could measure the light level continuously by measuring the speed at which the battery was charging. The harvester was capable of measuring this charging rate.The M3s developed for this study ran on only 40 nanowatts in standby mode, and 228 nanowatts when actively sensing. To give an idea of how miniscule this is, there are 1B nanowatts in a single watt. With the extremely small battery included in the M3, which can provide about only a few millions of a watt for a single hour, every nanoamp counts. The ability of the M3 to run on such low power was key to its success.Field work in Tahiti shows P. hyalina can take 10x more lightAfter local testing enabled by local Michigan snails, 50 M3s made it to Tahiti in 2017. Bick and Lee joined forces with Trevor Coote, a well-known conservation field biologist and specialist on the French Polynesian snails.The team glued the sensors directly to the rosy wolf snails, but P. hyalina is a protected species and required an indirect approach. They are nocturnal, typically sleeping during the day while attached underneath leaves. Using magnets, the team placed M3s both on the tops and undersides of leaves harboring the resting P. hyalina. At the end of each day, Lee wirelessly downloaded the data from each of the M3s.The data revealed a dramatic difference in how much sun reached the habitats of the surviving P. hyalina as opposed to the rosy wolf snail. During the noon hour, the P. hyalina habitat received on average 10 times more sunlight than the rosy wolf snails. Specifically, the average light intensity reached 7,674 to 9,072 lux for the P. hyalina habitat, but only 540 to 762 lux for the rosy wolf snail.The researchers suspect that the rosy wolf doesn’t venture far enough into the forest edge to catch P. hyalina, even under cover of darkness, because they wouldn’t be able to escape to shade before the sun became too hot.Model for future collaborations between biologists and engineersThe success of this project broke new ground from the perspective of the engineers as well as the biologists.“It’s underappreciated how large a step it is to go from the lab into the field and get meaningful data,” said Blaauw. “It’s essential to achieve success in order to propel the technology forward, but takes a great deal of trust among the collaborators.”“The M3 really opens up the window of what we can do with invertebrate behavioral ecology and we’re just at the foothills of those possibilities,” Ó Foighil added.This project has already facilitated a subsequent collaboration between engineering and ecology and evolutionary biology tracking monarch butterflies. And more projects are in the works.The research is published by Communications Biology in “Millimeter-sized smart sensors reveal that a solar refuge protects tree snail Partula hyalina from extirpation,” by Cindy Bick, Inhee Lee, Trevor Coote, Amanda Haponski, David Blaauw and Diarmaid Ó Foighil.The project was supported by U-M’s MCubed program, created to stimulate and support innovative research among interdisciplinary teams. Additional funding was provided by the Department of Ecology and Evolutionary Biology and by National Science Foundation and Arm Ltd. funding to the Blaauw lab.

Modeling a Circular Economy for Electronic Waste

Research, Electrical & Computer, Chemical & Petroleum, Banner

Think about how many different pieces of technology the average household has purchased in the last decade. Phones, TVs, computers, tablets, and game consoles don’t last forever, and repairing them is difficult and often as expensive as simply buying a replacement.Electronics are integral to modern society, but electronic waste (e-waste) presents a complex and growing challenge in the path toward a circular economy—a more sustainable economic system that focuses on recycling materials and minimizing waste. Adding to the global waste challenge is the prevalence of dishonest recycling practices by companies who claim to be recycling electronics but actually dispose of them by other means, such as in landfills or shipping the waste to other countries.New research from the Hypothetical Materials Lab at the University of Pittsburgh Swanson School of Engineering develops a framework to understand the choices a recycler has to make and the role that digital fraud prevention could have in preventing dishonest recycling practices. “Electronics have huge environmental impacts across their life cycle, from mining rare raw materials to the energy-intensive manufacturing, all the way to the complicated e-waste stream,” said Christopher Wilmer, the William Kepler Whiteford Faculty Fellow and associate professor of chemical and petroleum engineering, who leads the Hypothetical Materials Lab. “A circular economy model is well-suited to mitigating each of these impacts, but less than 40 percent of e-waste is currently estimated to be reused or recycled. If our technology is going to be sustainable, it’s important that we understand the barriers to e-waste recycling.”Some U.S. firms that have touted safe, ethical and green recycling practices never actually recycle much of what they receive; instead, their e-waste was illegally stockpiled, abandoned or exported. Between 2014 and 2016, the Basel Action Network used GPS trackers in electronics delivered to U.S. recyclers, showing that 30 percent of the products ended up overseas.The researchers developed a model framework that analyzes dishonest end-of-life electronics management and what leads recyclers to pursue fraudulent activities. They find that the primary way to ensure an e-waste recycler will engage in honest practices with minimum supervision is to make it the more profitable option, either by decreasing the costs of recycling or increasing the penalties for fraudulent practices. “The main barrier to honest recycling is its cost,” said lead author Daniel Salmon, a graduate student in the Department of Electrical and Computer Engineering. “One of our main findings is that if we find a way to make it more profitable for companies to recycle, we will have less dishonest recycling. Targeted subsidies, higher penalties for fraud and manufacturers ensuring their electronics are more easily recyclable are all things that could potentially solve this problem.”The researchers also suggest the use of the blockchain as neutral, third-party supervision to avoid fraudulent recycling practices.“Our model mentions the influence of monitoring and supervision, but self-reporting by companies enables dishonesty. On the other hand, something like the blockchain does not,” said Wilmer, who founded Ledger, the first peer-reviewed scholarly journal dedicated to blockchain and cryptocurrency. “Relying on an immutable record may be one solution to prevent fraud and align behaviors across recyclers toward a circular economy.”The work is part of a larger NSF-funded convergence research project on the circular economy, which is led by Melissa Bilec, deputy director of the Mascaro Center, associate professor of civil and environmental engineering, and Roberta A. Luxbacher Faculty Fellow at Pitt. The paper, “A Framework for Modeling Fraud in E-Waste Management,” (DOI: 10.1016/j.resconrec.2021.105613) was published in Resources, Conservation and Recycling and coauthored by Daniel Salmon and Christopher E. Wilmer at Pitt, and Callie W. Babbitt and Gregory A. Babbitt at Rochester Institute of Technology.