The most exciting work in artificial intelligence (AI) is happening in the field of machine learning using a technique known as deep learning. This process enables computers to learn by running massive amounts of data through a complex set of algorithms. The system is inspired by the biology of the human brain, and is called a neural network.
Building these systems requires a team of scientists and engineers who painstakingly design the set of rules and methods necessary to achieve the desired result. With the current AI boom underway, experts with the skills required to create these systems are in short supply, with only an estimated 10,000 people worldwide proficient in the task.
Hollywood vs. reality
Contrary to the depiction of AI in movies, each AI model is created for -- and can typically only accomplish -- a very specific and narrowly defined task. To use an extremely simplified example, if a system is designed to recognize cats, it can become extremely adept at distinguishing the furry felines -- but is completely incapable of identifying a dog. Consequently, scientists must create a new system for each subsequent task. These can include such diverse capabilities as the voice recognition used by smartphones or the language translation accomplished by Google Translate.
Back in May, Google debuted its AutoML project (ML is short for machine learning) which sought to automate the process of designing these machine learning systems. Creating an AI system that could build subsequent systems would shorten the amount of time necessary to deploy new models. In a blog post at the time, research scientists Quoc Le and Barret Zoph of the Google Brain team wrote:
In our approach (which we call "AutoML"), a controller neural net can propose a "child" model architecture, which can then be trained and evaluated for quality on a particular task. That feedback is then used to inform the controller how to improve its proposals for the next round. We repeat this process thousands of times -- generating new architectures, testing them, and giving that feedback to the controller to learn from.
The scientists noted that eventually the controller AI would become more proficient and achieve a greater degree of accuracy when training the child.
A child is born
The result of this research was an AI system dubbed NASNet, the AI child that was trained in computer vision and taught to detect objects, just like the technology deployed in self-driving cars. After the training process was repeated thousands of times, the system was tested against ImageNet image classification and COCO object detection data sets, "two of the most respected large scale academic datasets in computer vision."
When tasked to recognize objects -- like people and kites -- in real-time video streams, the child system achieved an 82.7% accuracy rating. This is more accurate than any state-of-the-art system currently available, and better than any built by humans. It also reduced the computational cost by half.
The future has arrived
Google believes that its latest research could be applied to many computer vision applications and has open-sourced NASNet. It believes that others will be able to build upon and improve these systems to "address multitudes of computer vision problems we have not yet imagined."
Currently, self-driving car technology is on the fast track to adoption, and improvements in computer vision systems will result in better safety for autonomous vehicles.