According to researchers at Google Brain, their newest artificial intelligence (AI) creation, AutoML, is not only capable of creating its own AIs, it is better at it than humans. The project, an early example of recursively self-improving AI, involved the use of reinforcement learning to automate the development of machine learning templates.
AutoML (“ML” is short for machine learning) acted as a controller neural network that spawned a child AI called NASNet. AutoML played the role of supervisor and teacher for its child AI, overseeing its ability to perform a specific task over and over again. In this case, NASNet was charged with real-time object detection, which it completed with record efficiency.
According to CEO Sundar Pichai, AutoML solves one of the most intractable problems for deep learning software engineers, which is selecting the best architecture for a neural network.
Google’s researchers say the development of computer vision algorithms will not only help to expand the field — which, by some estimates, has only 10,000 people worldwide with the ability to write such complex mathematical algorithms — but that it could also lead to huge improvements in self-driving cars and even enhanced assistance for visually impaired humans.
While recursively self-improving AI will lead to the exponential growth of AI, experts say that democratizing the field of AI by allowing non-experts to develop AI applications will lead to human growth as well.
In a blog post, Google Brain researchers wrote:
“We hope that the larger machine learning community will be able to build on these models to address multitudes of computer vision problems we have not yet imagined.”
In a time when anxiety over automated weapons and surveillance — which could both be dramatically strengthened by next-generation algorithmic AI — has reached fever pitch, perhaps more humans becoming engaged with AI is a good thing. With a “seed AI” such as AutoML taking care of the heavy-lifting, smart machines making new smart machines could help humans grow smarter in the process. It seems we’ve figured out the solution to the shortage of AI experts — create them.