Google AI self-made | 'sub-AI' | performance slightly more than man made

Earlier this year, Google announced that its AI system was able to invent its own encryption algorithm and generate its own AI, according to Google's official blog and Futurism, News Network recently reported that this AI created by the AI, the performance has defeated the AI ​​created by mankind: In the test, the "sub AI" system named NASNet achieved a correct rate of 82.7%, compared with the previously released similar AI products The result is 1.2% higher and the system efficiency is 4% higher.

In May 2017, researchers at Google Brain announced the development of Automated Artificial Intelligence (AutoML), a system that can produce its own "child AI." Recently, they decided to launch AutoML to date with the biggest challenge to date - - Try to use AI created by AutoML to defeat human design AI.

Team members automate the design of a machine learning model using a technique known as reinforcement learning. This time, AutoML's Identity is a controller neural network that develops a 'sub AI' for specific tasks. The newly generated 'child Called NASNet, identifies targets in the video, including humans, cars, traffic lights, handbags, backpacks, etc. AutoML, as a 'parent', assesses the performance of 'kids' NASNet and uses that information to improve 'child AI' , Then repeat this process thousands of times.

Team members tested 'Sub AI' NASNet on two datasets ImageNet (Computer Vision System Identification Project, currently the largest image recognition database in the world) and COCO Target Recognition, which they said is in the area of ​​computer vision The two most recognized large-scale academic datasets, on the order of magnitude, make testing very serious.

As a result, NASNet achieved 82.7% predictive accuracy on the validation set in the ImageNet test, 1.2% better than previously published results for the same type of artificial intelligence products and comparable with what was reported but unpublished on the preprinted Web site With 4% more system efficiency and 43.1% average accuracy for the largest models.Team members said NASNet will be used in a variety of applications that allow users to classify images and detect objects using the AI ​​system.

Chief editor punctuality

Robot can make robot, AI can design AI .It's no wonder to think about it, as long as the clear definition of the target, of course, a powerful computer faster than the human brain, sooner or later will replace people.But does not mean that AI can be independent of people's progress. AI is still tethered to the cage, and occasionally put into the track, run away nothing.When the AI ​​whim, set a goal for themselves, when it can be compared with people. Now far worse .