In order to give you a better service Deutsche Messe uses cookies. If you continue we assume that you consent to receive cookies on all Deutsche Messe websites. Legal Notes

Artificial Intelligence

Google’s AI writes better AI software than its developers

Google’s AutoML project was conceived as an AI that can be used to develop other AI systems. Now AutoML has produced more powerful and efficient systems than human developers.

02 Nov. 2017 Source: t3n Cornelia Dlugos

In May, Google presented its machine-learning project AutoML . The AI is intended to compensate for the shortfall of highly qualified technical personnel for AI programming and to create and improve new AI systems. The system performs thousands of simulations to find the parts of a code that can be improved, improves them, and then continues this process.

AutoML has now managed to surpass the human developers, as Wired reports. An AI produced by AutoML set a record for 82 percent accuracy in categorizing images based on content. For the task of marking the position of various objects in an image, the AI had a success rate of 43 percent, while the best system developed by humans achieved only 39 percent.

This means that Google’s AutoML project has delivered significant results, as even at Google only a few employees have the ability to develop such next-generation AI systems. Automating such processes will change the industry over the long term. "Currently they are still handcrafted by machine-learning experts, and there are only a few thousand scientists who are able to do that," says Google CEO Sundar Pichai, as quoted by Wired. "We want to make it possible for hundreds of thousands of developers."

Are AIs better at developing AIs?

For artificial intelligences, it is becoming easier to develop complex systems. But human intervention is by no means unnecessary. As the Chatbot Tay has demonstrated, for example, when unsupervised AI often builds connections to negative stereotypes, such as those based on ethnic or sex-specific identities. Human developers must therefore function as gatekeepers here.

While they require less time to construct AI systems, they require more resources for reviewing and refining them.