New Neural Network Slashes Sensor Data Overload
Modern technology collects huge quantities of data from sensors, with one estimate to drop global data from Internet of Things in about 73 olives (or 73 trillion GB) in 2025. With more data collection, the infrastructure required to store these data and operate it on this data as well.
But what if, instead of collecting all possible data from the sensor, we can be more selective, and only collect enough data to determine what we are looking for accurate? This is the approach proposed by researchers at Pennsylvania State University and the Massachusetts Institute of Technology. Their paper, which was recently published in Nature scientific reportsIt shows how the nervous network can achieve more than 90 percent accuracy with samples of 10 percent of the original sensor data.
“The way I see it, the edge computing will take a different direction because of what we did – or not only Edge, but also the edge used along with cloud computing,” says Sunsar Kumara, professor of industrial engineering in Pennsylvania and the author of the paper.
Inspires inspired by the human senses
Since the late 1980s, Kumara has been investigating the use of artificial intelligence of industrial systems, focusing on sensor data. Over the years, he has searched Forete transfers, wave theory, and chaos theory, among other ideas.
Recently, Ancor Verma, who was finishing his doctorate. In Pennsylvania, he came to Kumara at a new angle. “Humans can understand things with only a small amount of information,” says Verma. “The question we asked after that is, can we make machines to do the same?”
Attempts to address this problem face an obstacle in the theory of Nyquist-Shannon samples, a sporty guide that the signal sample rate should be twice the width of the frequency range to avoid avoiding information loss. To accurately measure the sound wave of Herz, for example, samples must be taken from 200 Hz (or more).
This theory indicates that a large volume of sensor data must be collected and processed for accurate results, which contributes to the increasing amount of growth from the sensor data collected and processed by modern devices.
A greatly effective nervous network
The researchers dealt with this problem with a “spectral deviation network of spectral transformation” or SIUN. It is a nervous network that uses “selective learning” to train on sensor data without using the total available data.
“We take samples at NYQUIST rates, but we do not collect every data point in this accuracy,” explains Verma, noting that SIUN depends on the random samples based on seeds to collect only part of the data. “It turns out that you can do this while keeping most of the data in the sign.” The researchers will be possible due to repetition often in sensory data groups.
The researchers tested SIUN for several data sets used to assess the discovery of breakdowns, such as the Case Western Reserve Data set, which includes a variety of data from good and loving ball bearings. The nerve network was asked to classify the bearings correctly as normal or wrong, and if they are wrong, determine the type of error.
The system was 96 percent accurate when only 30 percent samples of the initial data of this data set were taken. When tested against other data collections, SIUN was usually in the range of 80 to 90 percent minutes when less than 20 percent samples were taken from the initial data.
For comparison, the SIUN paper against a more traditional fodder nervous network (CNN) incited the ball -bearing data set. CNN won accurately, rating of errors with 99.77 percent, compared to 96 percent of Sion. However, CNN uses the entire data set and becomes a larger and more complex model. CNN contained more than 3 million teachers, while SIUN had less than 42000.
Simply: Sion’s efficiency overcame CNN, and it was not close. The researchers found that Swin “achieves a 435.01X reduction in the number of fluctuations required”, to classify the bearings within the data set. Although this is the best example of researchers, other tested data groups also found significant reinforcements, while reducing account requirements ranging from 8 times to 27 times compared to CNN.
New approach applications for sensor data
To bring point point home, I launch an inexpensive controlled and available Ferma. “We have published our program on Raspberry Pi Pico,” says Verma. PICO, which is $ 4, contains 264 kilograms of RAM and a 133 -megapixel dual -core processor. “It works on a few millionaires of power, but we are still able to infer this.”
Imagine that there are factories that produce things on Mars. We can not only buy additional graphics processing units on Mars. ” – Ankur Verma, Pennsylvania State
The researchers presented several patent applications related to technology and Verma, along with the co -author of Ayush Goyal, a company called Lightscline marketing the approach. They believe that the results of the paper may be related to many practical sensing tasks, but their lunar snapshot is literal. They want to take the idea to space.
“Imagine if we had settlements in space, or on Mars, and there are factories that produce things on Mars,” says Verma. “We can not only buy additional graphics processing units on Mars. We cannot put more cloud storage.”
While Mars factories may look great, the example represents the real world’s concerns. SMALLSAT RIDESHARE is the cost of launching a cost of $ 6000 per kilogram. At this rate, the setting of the NVIDIA DGX H200 in the orbit will cost more than $ 750,000. SIUN Mission’s approach can help accomplish space sensing tasks with lighter devices that are less expensive to launch them.
Kumara, the founder of Lightscline, also had a more example on the ground. He believes that SIUN can bring the advantages of artificial intelligence sensor to rural areas with less stars reaching artificial intelligence devices. “Imagine that even manufacturing sites in rural areas, on the edge, you can do this account.” He says: “They can reach deeper visions of their manufacture and quality.”
This story was updated on January 28, 2025 to add that Soundar Kumara is also the founder of Lightscline.
From your site articles
Related articles about the web
2025-01-27 12:00:00
