AI capable of automatically constructing a "neural network" realizing AI appears

ByChris Searle

AI (artificial intelligence) realized by deep learning such as automatic driving car, voice recognition, image analysis, etc. is gaining considerable attention as computer technology which greatly changes society. What is essential is to reproduce the structure of the human brain on a computer "neural networkNormally it is developed by researchers, engineers and others, but the Oak Ridge National Laboratory (ORNL) of the Department of Energy of the United States has developed an algorithm for AI to automatically generate a neural network We have succeeded in developing.

An AI for AI: New Algorithm Poised to Fuel Scientific Discovery | The Official NVIDIA Blog
https://blogs.nvidia.com/blog/2018/01/12/an-ai-for-ai-new-algorithm-poised-to-fuel-scientific-discovery/

ORNL is a supercomputer whose performance has been upgraded by GPUTitanWe have developed an algorithm to automatically generate a neural network using ". Titan can model the network of nerves in the human brain and perform the process of "learning deep learning (deep learning)".

It is named after the abbreviation "Multi-node Evolutionary Neural Networks for Deep Learning" (multi-node evolutionary neural network for deep learning)MENNDLis. It is a neural network that can be evaluated and tested by a scientist and recommended for a neural network, in other words, the computer itself can design a computer. Titan has succeeded in dramatically shortening processing time by using GPU acceleration and Steven Young, a research scientist who is a member of ORNL's Nature Inspired Machine Learning team said, "MENNDL is We will save money and make it possible to reach scientific discovery faster. "

Titan has 18,888 NVIDIA GPU accelerators "TeslaIt is equipped with MENNDL scaling by simultaneous processing, it is possible to predict what is most suitable for the output required by testing and thousands of networks simultaneously. MENNDL has already been used in many scientific fields such as elementary particles and space science, and we are starting to get results.

Introducing Titan | The World's # 1 Open Science Supercomputer


National Institute of High Energy Physics belonging to Ministry of EnergyFermi National Accelerator Research InstituteThen, it is one of elementary particlesNeutrinoWe are doing research on. Because it is very difficult to detect neutrinos in nature, the laboratory conducts research by looking at the response when irradiating high-altitude beams on objects.

At this time, the data obtained from various sensors becomes enormous, and researchers need to investigate this data in detail and grasp the phenomenon. Once it took a couple of months to do this work, MENNDL was able to complete the same process in just 24 hours. Mr. Young said, "MENNDL has been able to create a network that can be used" in only one day, while scientists have been struggling for several months with the deep learning framework so far. " I am talking.

ByKathryn Grim

The benefit that this MENNDL brings is to find useful information from a huge amount of dataData MiningIt will be most utilized in the field of. It seems that the evolution of computer technology / AI technology still does not know where to stay.