In May Google Brain team presented AutoML: a system, that automates the process of designing machine learning models. As for now, the designing of such model is a sophisticated process, which may take many hours of work for researchers and engineers. And in the first attempt, AutoML mainly improved the existing models and offered better parameters and settings or created models for relatively small neural networks.

But last month the team published a research, explaining how they applied AutoML to two of the most respected large-scale academic datasets in computer vision. And the results are exciting: the accuracy of prediction by automatically designed neural networks was surpassed all the published and yet to be published (found on arxiv.org) results at least by 1.2% while having 28% reduction in computational demand.

To put it simply, Google Brain’s method designed two connected neural networks: one of them controls the process and the other is learning, while receiving constant feedback from the first (left and right networks on the figure below, accordingly).

(Architecture of the networks, taken from the published paper)

The job of Machine Learning model designer has been replaced soon after appearing. And now I’m wondering, is there anything that human will still do better than machines let’s say 50 years from now?