Google calls it a means to make machine learning models more accessible. AutoML’s controller neural net can propose “child models” that can be further trained and evaluated. This child AI, called NASNet, can be used to recognize objects in real-time.
After repeating this training process thousands of time and improving NASNet, it was tested on the ImageNet image classification and COCO object detection sets; these are two most respected large-scale data sets in computer vision. The results were surprising as NASNet outperformed any other computer vision system, according to Google, as reported by Futurism.
As per the results on ImageNet image classification, NASNet surpassed previous models and showed a prediction accuracy of 82.7%. Moreover, it performs 1.2% better than all previously published results. The computational costs are also very low.
The state-of-the-art predictive performance on COCO object detection task, Google’s largest model achieved 43.1% mAP, which is 4% better than previously published state-of-the-art.
It’s also worth noting that Google has open-sourced NASNet for inference on image classification and detection in TensorFlow repos.
These types of huge advancements in the field of artificial intelligence and machine learning are the reasons we need more stringent regulations and better ethical standards for AI. What are your thoughts on this advancement? Share your views and become a part of the conversation.