Google’s AutoML utilizes reinforcement learning to build superior versions of itself.

 

Earlier this year (May 2017), the Google Brain Team unveiled their AutoML AI (Artificial Intelligence) platform capable of creating its own AIs. Moving forward, the Team recently announced that the same AI platform has succeeded in creating an offspring (or child) that has outperformed anything made by Google engineers. AutoML used deep learning to build the new AI, which uses multiple layers of neural networks to translate high-level abstractions, recognize patterns and comprehend varied concepts. More directly, it mimics human learning capabilities.

 

At its core, the AutoML project was created to make it easier to design machine learning models by automating the process. For example- a controller neural net proposes a child AI model, which is then trained and then evaluated for quality by performing a specific task. The resulting feedback is then used to inform the controller on how to improve that function for the next round of testing. This process is performed thousands of times over, generating new architectures for the controller to learn from.

 

An example of object detection and recognition utilizing NASNet with Faster-RCNN. (Image credit: Google)

 

This last November, AutoML was used to create NASNet- an offspring AI designed for object recognition, outperforming other AIs made for academic competitions. To test the offspring, Google’s Team applied it to both the ImageNet image classification and COCO (COmparing Continous Optimizers)  datasets, two of the best large-scale datasets in computer vision. With ImageNet, NASNet obtained a prediction accuracy of 82.7%- 1.2% better than any other AI object recognition platforms. As for COCO, the child AI garnered a 43.1% mAP (mean Average Precision), 4% better than other AIs that have undergone the predictive performance on the object detection task.

 

Considering computer-vision programs are in high demand, AutoML’s child NASNet could have far-reaching applications; including creating sophisticated AI-driven robots, increased autonomous vehicle object recognition and avoidance and even help visually impaired people gain/regain their sight.

 

Realizing that the new AI child could be reused for any number of computer vision applications, the Brain Team has open-sourced NASNet for inference on image classification and object detection in the Slim and Object Detection repositories on their GitHub page. Google hopes others will take advantage of their machine learning platform and build their own AI creations using the NASNet software.

 

Have a story tip? Message me at: cabe(at)element14(dot)com

http://twitter.com/Cabe_Atwell