In the spring of this year, Google Brain engineers presented AutoML artificial intelligence capable of creating its own unique AI without human intervention. Not so long ago it became known that AutoML was the first to create the NASNet computer vision system, significantly superior to all analogues created by man. This AI-based system can become a serious help in the development of, for example, autonomous cars, as well as in robotics, allowing you to bring the vision of robots to a whole new level.
AutoML develops according to the so-called reinforcement learning system. In fact, it is a control neural network that independently develops completely new neural networks for any specialized tasks. In this case, the main goal of AutoML was to create a system for the most accurate recognition of objects on video in real time. AI independently trained a new neural network, tracking its errors and making adjustments to its work. The learning process was repeated many thousands of times until the system became operational. Moreover, it surpassed all existing similar neural networks created and trained by humans.
According to the official statement of Google, the recognition accuracy of NASNet is 82.7%. This is 1.2% better than the previous record set in September this year by experts from Oxford and Momenta. The neural network also turned out to be 4% more efficient than analogs with 43.1% average accuracy. A simplified version of NASNet adapted for mobile platforms outperforms similar neural networks by more than 3%. In the future, this system can be used to create autonomous cars, because for them computer vision is incredibly important. In the meantime, AutoML continues to create new neural networks, and who knows what heights it will be able to achieve in the near future.
Sergey Gray