Google has released an advanced version of its AI algorithm for captioning pictures to developers.
The company already employs a stripped-off version of this AI for recognizing pictures in Google Photos. The full-powered algorithm is said to have a 93.9 percent accuracy in identifying the content of pictures. The new algorithm can also make use of human language to describe pictures.
According to Google Research Blog, Google Brain team's AI image captioning system is now improved and has better accuracy rating. In 2014, the system was able to achieve 89.6 percent accuracy rating by using the Inception V1 image classification model. However, the new system using Inception V3 model is capable of emulating natural human language and can even caption the pictures by describing actions and colors.
The Brain team elaborated that the improvements were brought about using a process where humans were asked to caption a large number of pictures. The data was then fed into TensorFlow. While the algorithm may reuse human-written description for appropriate pictures, it can also write its own captions.
Android Authority reported that Google released its updated TensorFlow to developers expecting that they will be able to put the technology to diverse uses. The latest iteration of TensorFlow is not only more accurate but is also faster to train. Google blog stated that time per training step taken by TensorFlow is 0.7 seconds, in comparison to 3 seconds taken by DistBelief using a Nvidia K20 GPU.
Google has also given the developers the option of using the Inception-ResNet-V2 model instead of the standard Inception V3 image classification model.