[ad_1]
MIT has developed a method to make compiling neural networks less time-consuming, faster and more cost-effective. This should make the IA more democratic and accessible. […]
Researchers at Mbadachusetts Institute of Technology (MIT) have developed a new method that makes computer science using neural networks not only more efficient, but also much faster. The experts hope to make the development of machine learning systems more cost-effective, while opening access to this technology to a wider audience.
As a comparative example, researchers attract Google. The search engine leader needed about 48,000 GPU hours to create a single complicated neural network for image clbadification. In contrast, the new MIT method could reduce this time to 200 hours GPU.
This is achieved by removing unnecessary components, the researchers said. In addition, some filters in the form of customizable settings are used. Each of these filters processes the pixels of the image in a raster network. In principle, the filters systematically summarize several pixels and thus compress the image.
Binarization level path
In addition, researchers use a system called "binarization at the path level". Normally, a neural network stores all path possibilities completely. However, this requires a lot of storage capacity. However, in the case of MIT members, only one of the scanned paths is recorded. By combining several times and badyzing all the paths originally created, the algorithm finally determines the one with the highest probability. All others are deleted. With this process, no neuron would be rejected, the "cut off" path, however, completely changes the entire network, the researchers said. However, this method does not affect the accuracy or the loss of data.
The huge cost and time savings not only help large companies such as Google, IBM and others to form artificial intelligence models, but also small businesses and developers who otherwise do not have the financial resources necessary to do it.
* Alexandra Lindner is a writer at PCTipp.
Source link
