Futurologists and science fiction authors have made it clear to us that the fear of artificial intelligence will become justified when it can create its own kind without human intervention. And it seems that this day has come.
Recently, Google Brain engineers introduced AutoML , a project that, itself being artificial intelligence, is capable of designing child AIs by generating small neural networks similar to those created by humans. The experiment showed that AutoML copes with this very well. Thus, the “robot” designed the NASNet computer vision system, which is superior to all existing analogues. The AI completed the task of object classification using ImageNET's large computer vision dataset . NASNet was faced with the task of recognizing objects in video in real time as accurately as possible. As it turned out, AutoML independently trained the child neural network, repeating the training process many thousands of times. AutoML artificial intelligence identified errors in NASNet operation, processed them and made changes to avoid them in the future.
When engineers tested NASNet on ImageNet and COCO image sets , it outperformed all existing computer vision systems. The accuracy of image prediction in the control set was 82.7%, which is 1.2% higher than the previous indicator of the Inception family. In addition, the neural network turned out to be 4% more effective than its analogues with 43.1% average accuracy, and the version adapted for mobile platforms with 74%. The efficiency of the daughter system could be used in autonomous cars or in creating robots for visually impaired people. NASNet is open source and can be found in the Slim and Object Detection repositories for TensorFlow.
GO TO FULL VERSION