Google uses AutoML, or automated algorithm design, to build models that can automatically classify tons of images and detect objects -- processes that are very useful in search. In a blog posted earlier this week, Google researchers explain the automation of machine-learning algorithms based on large sets of data to identify objects in photographs, for example.
The researchers found that using their approach, AutoML can design small neural networks that perform in a similar way to neural networks designed by humans, and the results were compressed to small datasets. But the researchers wanted to know how the method would perform on larger, more challenging datasets. People have invested in these larger systems, but what about machines?
Tweaking the method, Google researchers redesigned the search space so that AutoML could find the best stackable layer to create a network.
he algorithm actually can identify images in a photograph. For example, take an image of people walking on the sand near the ocean. The algorithm identified kites flying in the sky over the ocean and the people walking on the beach by the water.
Researchers say the "largest model achieves 43.1% mAP, which is 4% better than the previous, published state-of-the-art."
Marketers should care because the algorithm enables a higher percentage of accuracy when it comes to identifying objects. (You can read more about it here.)
Yitaek Hwang, director of research and development at Leverege, earlier this year wrote about the promises versus the reality of AutoML. He identifies a Boston-based software company called DataRobot that is working to automate machine learning solutions. In other words the company automates the development of artificial intelligence (A.I.).
Hwang writes about the usefulness of the technology and the future of AutoML. It will partly enable the ability of machines to remember what they learned and apply it to new situations. That's the automated part of ML.