Deep Learning Computer systems are a system based on Artificial Neural Networks that mimic human learning methods and are now a hot topic in the field of computer science. In-depth learning has not only helped the development of facial and voice recognition software, but also brought a wealth of data-aided diagnostics to the medical field.
However, the computations that these systems need to perform are very complex and rigorous, and computing is still a challenge, even for the most configured computers.
At present, researchers have developed a new type of calculation - they use photons instead of electrons as the information delivery medium, which greatly improves the speed and efficiency of some deep learning computations. The result was done by 11 people, including Yichen Shen, a postdoctoral fellow at MIT, graduate student Nicholas Harris, professors Marin Soljacic and Dirk Englund, and researchers from elsewhere. At present, their papers have been published in Nature Photonics.
Professor Soljacic said that optical computers have been advocated by many researchers for many years because people over-hyped it before they seemed a bit disappointing and contrary to expectations. Although many people advocate the use of optical computers, the research results are not practical. The optical neural network system studied by this team can be applied to many programs.
The architecture of a traditional computer is not very efficient in the computation required to handle some important neural network tasks. Such tasks often involve repeated matrix multiplications that require intensive calculations on the CPU and GPU.
After years of research, the MIT proposed an opposite way to perform this operation. "Once you debug this chip, it can run matrix multiplication, which in principle requires very little power," Soljacic said. "We've already demonstrated the key modules, but they have not yet been fully released."
Soljacic explains the new system in analogy. Even ordinary ophthalmic lenses undergo complex calculations (also known as Fourier transforms) over the light waves that pass through it. The way the beam is calculated in a new photonic chip is similar to its basic principle. This new method uses multiple beams to propagate and their waves interact to create an interference pattern that conveys the result of the expected operation. Researchers call the device a programmable nanophotonic processor.
According to Yichen Shen, using this architecture of optical chips, in principle, can be carried out in the traditional artificial intelligence algorithms, much faster than traditional electronic chips, but using less than one-thousandth of energy. "The natural advantage of running matrix multiplication with light is that it plays an important role in acceleration and power savings, because intensive matrix multiplication is the most power-hungry and time-consuming part of AI algorithms," he said.
The new programmable nanophotonic processor was developed by Harris and co-workers at Englund Labs. It uses a series of waveguides (electromagnetic wave devices that transmit microwave bands) that are interconnected and can be modified as needed to program specific calculations. "You can program in any matrix operation," Harris said. The processor guides the light through a series of coupled photonic waveguides. They demanded that the staggered layers of the device apply a non-linear activation function that would be analogous to the actions of the brain's neurons.
To demonstrate this concept, the team set up a programmable nano-photon processor to implement a neural network that recognizes four basic vowels. Even with this basic system, they achieve 77% accuracy, compared to 90% for traditional systems. Soljacic said there is "no material barrier" to enhancing the accuracy of the system.
Englund added that programmable nanophotonic processors also have other applications, including signal processing for data transmission. "High-speed analog signal processing is faster than other methods of converting signals into digital form because light is a natural analog medium," he said. "This method can be handled directly in the analog domain."
The team said they also need more effort and time to make the system more useful. Once this system has been enhanced and fully functional, many user stories can be found, such as data centers or security systems. Harris said the system can also benefit unmanned vehicles or unmanned aerial vehicles and is suitable for situations where "a lot of calculations are needed, but you do not have much power and time."
Hardfacing Pipe is produced through an oscillating table with a welding head to arc weld the chromium carbide into the inside of the pipe. We can handle diameters from 6" to 24" and then we manufacture pipes from our chrome-plated sheets. The cover layer is deposited perpendicular to the material flow through the pipe, resulting in a longer wear life
Typical materials for cladding in common use at HP Welding include HP700, HP700S, HP750X, HP750P and stainless steels, as well as many less common alloys.
With over 30 years of experience on the production of hardfacing products, and the technical expertise to back them up, HP Welding can meet any base material and cladding combination.
The advantages of HP Hardfacing Pipe
1. Flat surface and nice appearance
2. Good weldability to others working pieces
3. Excellent wear resistance
HP Hardfacing Pipe can be widely applied to mining site, cement plant, metallurgical equipment, dredging and power generation and any other fields where need to combat with abrasion.
Alloy Steel Pipe,Hardfacing Iron Pipe,Chrome Alloy Hardfacing Super Wear Resistance Pipe
HuiFeng Wear Resistant Group , https://www.hpwearsolution.com