Monday, January 22nd, 2018

Time	Event
10:59a	Engineers design artificial synapse for “brain-on-a-chip” hardware When it comes to processing power, the human brain just can’t be beat. Packed within the squishy, football-sized organ are somewhere around 100 billion neurons. At any given moment, a single neuron can relay instructions to thousands of other neurons via synapses — the spaces between neurons, across which neurotransmitters are exchanged. There are more than 100 trillion synapses that mediate neuron signaling in the brain, strengthening some connections while pruning others, in a process that enables the brain to recognize patterns, remember facts, and carry out other learning tasks, at lightning speeds. Researchers in the emerging field of “neuromorphic computing” have attempted to design computer chips that work like the human brain. Instead of carrying out computations based on binary, on/off signaling, like digital chips do today, the elements of a “brain on a chip” would work in an analog fashion, exchanging a gradient of signals, or “weights,” much like neurons that activate in various ways depending on the type and number of ions that flow across a synapse. In this way, small neuromorphic chips could, like the brain, efficiently process millions of streams of parallel computations that are currently only possible with large banks of supercomputers. But one significant hangup on the way to such portable artificial intelligence has been the neural synapse, which has been particularly tricky to reproduce in hardware. Now engineers at MIT have designed an artificial synapse in such a way that they can precisely control the strength of an electric current flowing across it, similar to the way ions flow between neurons. The team has built a small chip with artificial synapses, made from silicon germanium. In simulations, the researchers found that the chip and its synapses could be used to recognize samples of handwriting, with 95 percent accuracy. The design, published today in the journal Nature Materials, is a major step toward building portable, low-power neuromorphic chips for use in pattern recognition and other learning tasks. The research was led by Jeehwan Kim, the Class of 1947 Career Development Assistant Professor in the departments of Mechanical Engineering and Materials Science and Engineering, and a principal investigator in MIT’s Research Laboratory of Electronics and Microsystems Technology Laboratories. His co-authors are Shinhyun Choi (first author), Scott Tan (co-first author), Zefan Li, Yunjo Kim, Chanyeol Choi, and Hanwool Yeon of MIT, along with Pai-Yu Chen and Shimeng Yu of Arizona State University. Too many paths Most neuromorphic chip designs attempt to emulate the synaptic connection between neurons using two conductive layers separated by a “switching medium,” or synapse-like space. When a voltage is applied, ions should move in the switching medium to create conductive filaments, similarly to how the “weight” of a synapse changes. But it’s been difficult to control the flow of ions in existing designs. Kim says that’s because most switching mediums, made of amorphous materials, have unlimited possible paths through which ions can travel — a bit like Pachinko, a mechanical arcade game that funnels small steel balls down through a series of pins and levers, which act to either divert or direct the balls out of the machine. Like Pachinko, existing switching mediums contain multiple paths that make it difficult to predict where ions will make it through. Kim says that can create unwanted nonuniformity in a synapse’s performance. “Once you apply some voltage to represent some data with your artificial neuron, you have to erase and be able to write it again in the exact same way,” Kim says. “But in an amorphous solid, when you write again, the ions go in different directions because there are lots of defects. This stream is changing, and it’s hard to control. That’s the biggest problem — nonuniformity of the artificial synapse.” A perfect mismatch Instead of using amorphous materials as an artificial synapse, Kim and his colleagues looked to single-crystalline silicon, a defect-free conducting material made from atoms arranged in a continuously ordered alignment. The team sought to create a precise, one-dimensional line defect, or dislocation, through the silicon, through which ions could predictably flow. To do so, the researchers started with a wafer of silicon, resembling, at microscopic resolution, a chicken-wire pattern. They then grew a similar pattern of silicon germanium — a material also used commonly in transistors — on top of the silicon wafer. Silicon germanium’s lattice is slightly larger than that of silicon, and Kim found that together, the two perfectly mismatched materials can form a funnel-like dislocation, creating a single path through which ions can flow.  The researchers fabricated a neuromorphic chip consisting of artificial synapses made from silicon germanium, each synapse measuring about 25 nanometers across. They applied voltage to each synapse and found that all synapses exhibited more or less the same current, or flow of ions, with about a 4 percent variation between synapses — a much more uniform performance compared with synapses made from amorphous material. They also tested a single synapse over multiple trials, applying the same voltage over 700 cycles, and found the synapse exhibited the same current, with just 1 percent variation from cycle to cycle. “This is the most uniform device we could achieve, which is the key to demonstrating artificial neural networks,” Kim says. Writing, recognized As a final test, Kim’s team explored how its device would perform if it were to carry out actual learning tasks — specifically, recognizing samples of handwriting, which researchers consider to be a first practical test for neuromorphic chips. Such chips would consist of “input/hidden/output neurons,” each connected to other “neurons” via filament-based artificial synapses. Scientists believe such stacks of neural nets can be made to “learn.” For instance, when fed an input that is a handwritten ‘1,’ with an output that labels it as ‘1,’ certain output neurons will be activated by input neurons and weights from an artificial synapse. When more examples of handwritten ‘1s’ are fed into the same chip, the same output neurons may be activated when they sense similar features between different samples of the same letter, thus “learning” in a fashion similar to what the brain does. Kim and his colleagues ran a computer simulation of an artificial neural network consisting of three sheets of neural layers connected via two layers of artificial synapses, the properties of which they based on measurements from their actual neuromorphic chip. They fed into their simulation tens of thousands of samples from a handwritten recognition dataset commonly used by neuromorphic designers, and found that their neural network hardware recognized handwritten samples 95 percent of the time, compared to the 97 percent accuracy of existing software algorithms. The team is in the process of fabricating a working neuromorphic chip that can carry out handwriting-recognition tasks, not in simulation but in reality. Looking beyond handwriting, Kim says the team’s artificial synapse design will enable much smaller, portable neural network devices that can perform complex computations that currently are only possible with large supercomputers. “Ultimately we want a chip as big as a fingernail to replace one big supercomputer,” Kim says. “This opens a stepping stone to produce real artificial hardware.” This research was supported in part by the National Science Foundation. (Comment on this)
10:59a	A new approach to rechargeable batteries A type of battery first invented nearly five decades ago could catapult to the forefront of energy storage technologies, thanks to a new finding by researchers at MIT. The battery, based on electrodes made of sodium and nickel chloride and using a new type of metal mesh membrane, could be used for grid-scale installations to make intermittent power sources such as wind and solar capable of delivering reliable baseload electricity. The findings are being reported today in the journal Nature Energy, by a team led by MIT professor Donald Sadoway, postdocs Huayi Yin and Brice Chung, and four others. Although the basic battery chemistry the team used, based on a liquid sodium electrode material, was first described in 1968, the concept never caught on as a practical approach because of one significant drawback: It required the use of a thin membrane to separate its molten components, and the only known material with the needed properties for that membrane was a brittle and fragile ceramic. These paper-thin membranes made the batteries too easily damaged in real-world operating conditions, so apart from a few specialized industrial applications, the system has never been widely implemented. But Sadoway and his team took a different approach, realizing that the functions of that membrane could instead be performed by a specially coated metal mesh, a much stronger and more flexible material that could stand up to the rigors of use in industrial-scale storage systems. “I consider this a breakthrough,” Sadoway says, because for the first time in five decades, this type of battery — whose advantages include cheap, abundant raw materials, very safe operational characteristics, and an ability to go through many charge-discharge cycles without degradation — could finally become practical. While some companies have continued to make liquid-sodium batteries for specialized uses, “the cost was kept high because of the fragility of the ceramic membranes,” says Sadoway, the John F. Elliott Professor of Materials Chemistry. “Nobody’s really been able to make that process work,” including GE, which spent nearly 10 years working on the technology before abandoning the project. As Sadoway and his team explored various options for the different components in a molten-metal-based battery, they were surprised by the results of one of their tests using lead compounds. “We opened the cell and found droplets” inside the test chamber, which “would have to have been droplets of molten lead,” he says. But instead of acting as a membrane, as expected, the compound material “was acting as an electrode,” actively taking part in the battery’s electrochemical reaction. “That really opened our eyes to a completely different technology,” he says. The membrane had performed its role — selectively allowing certain molecules to pass through while blocking others — in an entirely different way, using its electrical properties rather than the typical mechanical sorting based on the sizes of pores in the material. In the end, after experimenting with various compounds, the team found that an ordinary steel mesh coated with a solution of titanium nitride could perform all the functions of the previously used ceramic membranes, but without the brittleness and fragility. The results could make possible a whole family of inexpensive and durable materials practical for large-scale rechargeable batteries. The use of the new type of membrane can be applied to a wide variety of molten-electrode battery chemistries, he says, and opens up new avenues for battery design. “The fact that you can build a sodium-sulfur type of battery, or a sodium/nickel-chloride type of battery, without resorting to the use of fragile, brittle ceramic — that changes everything,” he says. The work could lead to inexpensive batteries large enough to make intermittent, renewable power sources practical for grid-scale storage, and the same underlying technology could have other applications as well, such as for some kinds of metal production, Sadoway says. Sadoway cautions that such batteries would not be suitable for some major uses, such as cars or phones. Their strong point is in large, fixed installations where cost is paramount, but size and weight are not, such as utility-scale load leveling. In those applications, inexpensive battery technology could potentially enable a much greater percentage of intermittent renewable energy sources to take the place of baseload, always-available power sources, which are now dominated by fossil fuels. The research team also included Fei Chen at Wuhan University in China, MIT research scientist Takanari Ouchi, and postdocs Ji Zhao and Nobuyuki Tanaka. The work was supported by the French oil company Total S.A. (Comment on this)
11:59p	Location detection when GPS doesn’t work With billions of GPS devices in use today, people are beginning to take it for granted that services on their handheld devices will be location-aware. But GPS doesn’t work well indoors, and it’s not precise enough for several potentially useful applications, such as locating medical equipment in hospitals or pallets of goods in warehouses, or helping emergency responders navigate unfamiliar buildings. Professor of aeronautics and astronautics Moe Win has spent the last decade investigating the theory and practice of using wireless signals to gauge location. In 2010, his group published a series of papers deriving fundamental limits on the accuracy of systems that infer wireless transmitters’ locations based on features of their signals, such as power, angle of arrival, and time of flight. In the February issue of the journal IEEE Transactions on Information Theory, Win and two colleagues — Wenhan Dai, an MIT graduate student in aeronautics and astronautics, and Yuan Shen, an associate professor of electronic engineering at Tsinghua University, who did his graduate work at MIT — expand on those results. First, they show how changing a wireless localization system’s parameters — such as the power, bandwidth, and duration of its transmissions — alters the fundamental limits on its accuracy. This, in turn, allows them to determine the system configuration that yields the most accurate location inferences. They also provide practical localization algorithms that can approach those limits in real-world scenarios. “We are developing a theory to determine the fundamental limits of location inference within different sets of constraints,” Win says. “In other words, what’s the best we can do with given resources? Based on the theory, we develop algorithms that approach these limits, and then we go into experimentation. The fact that we have the goal of going to experimentation means that the algorithms have to be as efficient as possible.” Geometrical thinking The researchers’ theoretical approach assumes that the localization network consists of nodes with known positions, referred to as “anchors,” and nodes with unknown positions, referred to as “agents.” Wi-Fi access points distributed through an office building, for instance, could serve as anchors. Smartphones trying to determine their positions relative to the anchors would count as agents. Within the theoretical framework, the goal is something the researchers call “node prioritization” — that is, determining which of the available anchors should transmit, at what power and with what range of frequencies and signal durations, in order to achieve a balance between localization accuracy and consumption of system resources. A solution that produced very accurate measurements by allowing an anchor to blast so loud and long that no other communication over the network was possible, for instance, would not be considered optimal. The researchers’ theoretical analysis shows that the ability to adjust system parameters can consistently reduce localization error by 30 to 50 percent. The key to the new paper is a geometric interpretation of the problem of choosing and configuring anchors. The metric that the researchers use to assess the accuracy of location inferences depends on three different characteristics of the location information extracted from wireless signals. As such, it defines a three-dimensional mathematical space, which turns out to be bullet-shaped. The possible settings of all the anchors in the network also define a mathematical space, which is typically much larger. If the network has 20 anchors, then the corresponding settings define a 20-dimensional space. Win, Dai, and Shen, however, found a way to transform the high-dimensional space into a three-dimensional one: a polyhedron that represents all possible anchor configurations that meet certain resource constraints. Transposing both sets of data into the same three-dimensional space makes calculating the solution to the node prioritization problem much simpler and faster. The problem becomes finding the bullet — a representation of the localization error metric — that intersects the polyhedron at exactly one point. This point represents the network configuration that will provide the most accurate location inference. If the bullet and the polyhedron don’t intersect at all, then the error measurement is unachievable. If they overlap, then the error measurement is not as low as it could be. Once the point of intersection has been identified, it can be mapped back on to the higher-dimensional space, where it represents particular anchor settings. Close enough This method is particularly effective if the network’s resource constraints — in terms of transmission power, bandwidth, and duration — are treated as a single cumulative value. In this case, finding the point of intersection between the polyhedron and the bullet is computationally practical. In real-world circumstances, however, the limitations of the nodes may need to be considered individually. In that case, the shape of the polyhedron becomes more complex, and finding the point of intersection becomes more time consuming. To address this scenario, Win, Dai, and Shen also present an approximate algorithm for network configuration with individually constrained devices. In the paper, they were able to show that, while the approximate algorithm is much faster, its results are virtually indistinguishable from those of the full-blown optimization algorithm. (Comment on this)

<< Previous Day

2018/01/22 [Calendar]

Next Day >>