Power/Performance Bits: Jan. 8

Ferrimagnetic memory; ML without negative examples; moving 2D sensors.


Ferrimagnetic memory
Engineers at the National University of Singapore, Toyota Technological Institute, and Korea University propose a new type of spintronic memory that is 20 times more efficient and 10 times more stable than commercial ones.

In spintronic devices, data is stored depending on up or down magnetic states. Current devices based on ferromagnets, however, suffer from a few issues. “Ferromagnet-based memories cannot be grown beyond a few nanometers thick as their writing efficiency decays exponentially with increasing thickness. This thickness range is insufficient to ensure the stability of stored digital data against normal temperature variations,” explained Dr Yu Jiawei, who worked on the project during doctoral studies at NUS.

Instead, the team turned to ferrimagnets, which can be grown 10 times thicker without compromising on the overall data writing efficiency. “The spin of the current carrying electrons, which basically represents the data you want to write, experiences minimal resistance in ferrimagnets,” noted Rahul Mishra, a doctoral candidate at NUS.

A team led by Assoc Prof Yang Hyunsoo (second from left) from NUS Engineering has discovered that ferrimagnet devices can manipulate digital information more efficiently and with more stability than commercial spintronic digital memories. (Source: National University of Singapore)

“In ferrimagnets, the neighbouring atomic magnets are opposite to each other,” said Yang Hyunsoo, associate professor at NUS. “The disturbance caused by one atom to an incoming spin is compensated by the next one, and as a result information travels faster and further with less power. We hope that the computing and storage industry can take advantage of our invention to improve the performance and data retention capabilities of emerging spin memories.”

Using an electronic current, the researchers were able to write information in a ferrimagnet memory element which was 10 times more stable and 20 times more efficient than a ferromagnet.

Next, the team plans to look into the data writing and reading speed of their device. In addition, they are also planning to collaborate with industry partners to accelerate the commercial development.

ML without negative examples
Researchers at RIKEN proposed a new method of machine learning classification called Pconf that relies only on positive examples of the object targeted for identification, with no negative examples needed.

The researchers offer several scenarios where negative data is limited or missing, such as retailers who may have data on shoppers who made a purchase but not those who didn’t, or app developers trying to predict which users will continue using the app but are required to delete data from users who unsubscribe from the app.

“Previous classification methods could not cope with the situation where negative data were not available, but we have made it possible for computers to learn with only positive data, as long as we have a confidence score for our positive data, constructed from information such as buying intention or the active rate of app users. Using our new method, we can let computers learn a classifier only from positive data equipped with confidence,” said Takashi Ishida of RIKEN AIP.

The confidence score, which corresponds to the probability whether the data belongs to a positive class or not, is a vital part of the scheme and allows the system to learn a classification boundary only from positive data and information on its confidence.

Illustrations of the Pconf classification (on right) and other related classification settings. Red points are positive data, blue points are negative data, and gray points are unlabeled data. The dark/light red colors on the rightmost figure show high/low confidence values for positive data. (Source: Takashi Ishida, Gang Niu, Masashi Sugiyama)

To test the system, the team used images from two different labeled datasets where one object was chosen as the positive class. From the Fashion-MNIST dataset, “T-shirt” was chosen as the positive class and another item (such as “sandal”) as the negative class. Confidence scores were attached to the “T-shirt” images. They found that, in some cases, the method that didn’t access the negative data (e.g., “sandal” photos) was just as good as a method that involves using positive and negative data.

According to Ishida, “This discovery could expand the range of applications where classification technology can be used. Even in fields where machine learning has been actively used, our classification technology could be used in new situations where only positive data can be gathered due to data regulation or business constraints. In the near future, we hope to put our technology to use in various research fields, such as natural language processing, computer vision, robotics, and bioinformatics.”

2D sensor transfer
Engineers at Rice University developed a method to transfer flexible 2D sensors from the fabrication platform to curved and other smooth surfaces. Such sensors could monitor optical fibers to identify potential performance issues.

The team created a 10nm-thick indium selenide photodetector with gold electrodes and placed it onto an optical fiber. Because it was so close, the near-field sensor effectively coupled with an evanescent field – the oscillating electromagnetic wave that rides the surface of the fiber – and accurately detected the flow of information inside. The sensor added no weight and didn’t impede signal flow.

“This paper proposes several interesting possibilities for applying 2D devices in real applications,” said Jun Lou, a professor of materials science and nanoengineering at Rice. “For example, optical fibers at the bottom of the ocean are thousands of miles long, and if there’s a problem, it’s hard to know where it occurred. If you have these sensors at different locations, you can sense the damage to the fiber.”

Rice engineers have developed a method to transfer complete, flexible, two-dimensional circuits from their fabrication platforms to curved and other smooth surfaces. Such circuits are able to couple with near-field electromagnetic waves and offer next-generation sensing for optical fibers and other applications. (Source: Zehua Jin / Rice University)

While the process of transferring 2D materials from one surface to another has improved, doing so when electrodes or other components are involved is much more challenging. “Think about a transistor,” said Lou. “It has source, drain and gate electrodes and a dielectric (insulator) on top, and all of these have to be transferred intact. That’s a very big challenge, because all of those materials are different.”

To overcome the problem, the team used a sacrificial layer of polydimethylglutarimide (PMGI) underneath the circuit to be moved. The PMGI is then etched away before transfer. On top, they used a layer of polymethyl methacrylate (PMMA), a common way to move 2D materials. The team tested the method with materials including molybdenum diselenide and believe it will work for any 2D material.

While the team has so far only developed passive sensors, they say the technique will make active sensors or devices possible for telecommunication, biosensing, plasmonics, and other applications.

Leave a Reply

(Note: This name will be displayed publicly)