It seems that when it comes to artificial intelligence, Google is unmatched - the group has completely dominated the AI segment and today shares another impressive achievement with us.
Researchers from the Google Brain team have just unveiled an artificial intelligence system called AutoML that can generate its own AI - a "child" called NASNet, and far surpasses the capabilities of any competitor that gets out of the human hand. Researchers have moved away from analyzing words and images, and now artificial intelligence can already recognize objects such as people, cars, handbags and signal lights at intersections.
AI acquired these skills through reinforcement learning - AutoML plays the role of a “parent” here, who creates “children” for specific tasks, and then receives the appropriate feedback. This allows the AI to make adjustments and prepare for the next round of learning - it creates new architectures, validates them, collects comments, and the whole process is repeated thousands of times.
Google created NASNet as an object detection system, so AI had to pass appropriate accuracy tests on huge databases. ImageNet and COCO are considered two key systems with which the scientists' new work has gone well, surpassing all currently available computer image recognition solutions. NASNet hit 82.7%, 1.2% higher than the previous leader, and 4% more efficient, with a 43.4% scaling factor (medium precision).
According to scientists, NASNet can be useful in many different applications and make its solution available open source, awaiting feedback that will help further work on image classification and object detection. As the developers say: - We hope that the machine learning community can use our model to discuss problems with pattern recognition that we don't even know about.
Of course, as always in such situations, there are already voices of the opposition and doubts indicating the unethical decision. Because if an AI is able to create its "children" much faster than humans, it can never keep up with it, and there is a risk that it will transfer its prejudices, etc., to the creations. In short, it is time to develop some AI standards and think about security issues so that such advanced tools do not fall into the wrong hands, in which AI can become very dangerous.
It seems that when it comes to artificial intelligence, Google is unmatched - the group has completely dominated the AI segment and today shares another impressive achievement with us. Researchers from the Google Brain team have just unveiled an artificial intelligence system called AutoML that can generate its own AI - a "child" called NASNet and significantly.
Pavel Zaitsev
Promotional video: