Google Inc.'s improved artificial intelligence powered algorithm can better recognize and classify photos and may achieve "human-level" accuracy soon, the company's engineer said Wednesday.
In November 2016, Google announced that it has applied a machine learning algorithm to its computer vision model of Google Photos to produce captions that can accurately describe images. The service -- both available online and through mobile applications -- can back up, organize and label the photos automatically so that the users can quickly find them.
Neil Alldrin, an engineer from Google Inc., talks to reporters in Seoul through a video conference on March 22, 2017. (Yonhap)
"We hope the computer vision model goes beyond average human level," Neil Alldrin, an engineer at Google's image search team, told reporters in Seoul through a video conference.
Google currently uses open-source data for its "Inception" model that is better at picking out individual objects in a single image.
The AI was first trained in 2014 and has steadily improved since, Google said, adding the improved system is faster to train and produces more detailed descriptions.
The most recent version of the system uses the third version of Inception, which undergoes a fine-tuning phase in which its language components are trained through human-generated captions.
"Some issues still come up. That's why we try to closely work with users," Alldrin said, adding that Google aims to work with users and researchers to build a better database that currently has over 9 million images.
"Machine learning has accelerated the development of vision systems, has improved performance and scales with more data," he said.
Google said the AI may someday understand the context and deeper meaning of a scene. (Yonhap)
