Google has announced the next big step in artificial intelligence development with a new approach to machine learning that can be used to use neural networks to create even more efficient neural networks. Basically, we are talking about teaching a machine to create its own kind.
Artificial neural networks are being designed to mimic the learning process of the brain, and according to Google, its new technology, called AutoML, has the potential to make these networks even more powerful, more efficient and easier to use.
Google CEO Sundar Pichai set an example of how AutoML works by speaking at Google I / O 2017, an annual event for hardware and software developers where the company typically presents or at least talks about the products it is currently working on.
“It works like this: we take a set of candidates for neural networks - let's call them baby neural networks - and repeatedly run a ready-made neural network through them to find errors until we get an even more efficient neural network”, - said Pichai.
This process is called stimulated learning, where a reward will be given to the computer for finding bugs. By the same principle, for example, they teach new tricks to dogs. Of course, in the case of computers, this requires enormous computing power, but the power of Google's equipment has already reached such a level that one neural network can easily analyze the work of another neural network.
It takes a real team of computer engineering experts and a huge amount of time to create a neural network, but thanks to AutoML, in the future, almost any user will be able to build their own AI system and program it to perform absolutely any task.
“We hope that AutoML technology, which is currently only available to a few research centers, in three to five years will become available to hundreds, and better thousands of neural network developers who want to use them for their specific purposes,” wrote Pichai in the official blog.
The scheme of the AutoML technology: multi-level analysis of the operation of neural networks to determine the most intelligent of them
Promotional video:
Machine learning - an attempt to give a computer the ability to draw its own conclusions based on available information - is just one of the approaches in the development of artificial intelligence, which includes two important aspects: the learning process and the actual ability to independently draw conclusions based on it. With training, everything is relatively clear. Show the computer a hundred thousand pictures of cats and dogs, and it will eventually figure out what combination of pixels each of these animals makes. The second part is a little more complicated. After all, it is here that the machine is required to show what it has learned, and on the basis of this learning, independently come to a logical guess. Make a conclusion.
Now replace cats and dogs with neural networks, and you will get an idea of how AutoML works, which, instead of recognizing animals, recognizes which of the presented systems is the most intelligent. According to Google, even now the level of AutoML is already such that it can be more effective than human experts in finding the best approaches to solve specific problems. In the future, this will significantly simplify the process of creating new AI systems, since in fact they will be created by their own kind.
AutoML is still in its early stages at the moment, Google says, but AI, machine learning and deep machine learning (advanced machine learning methods based on simulating the work of neurons in the human brain) are all already finding their way in one way or another. in those applications and areas that we use and in which we find ourselves on a daily basis.
In a demonstration on stage at the I / O conference, Google engineers showed how their machine learning technology can significantly brighten very dark images or, for example, remove various noise from them. And all these actions the machine is able to perform only relying on information obtained through the analysis of millions of other clear samples of images. Google notes that their supercomputers have now become more efficient than humans in the process of recognizing what is in the photo. Based on this technology, a custom Google Lens application will soon be released, which can effectively determine which flower (or flowers) is in front of you (or in the pictures) through the smartphone's camera.
In the future, such super-powerful algorithms based on deep learning will definitely find a place for their application in medicine, where systems based on them will detect signs of malignant tumors in images and in most cases do this much more efficiently than professional surgeons.
With AutoML technology, AI platforms will learn faster and be much smarter. True, this moment will have to wait a little longer than the release of the promised "flower application" for the Android platform. However, up to this point, application developers and scientists will have plenty of time to get to know AutoML better.
“We think that this technology will lead to the emergence of new neural networks and the opening up of opportunities where even non-experts will be able to create their own personal neural networks for their specific needs, which, in turn, will only increase the ability of machine learning technologies to exert more influence on us all. - say Google scientists Kuok Le and Barrett Zof.
NIKOLAY KHIZHNYAK