An important step towards the creation of artificial intelligence
Engineers have taught the brain from artificial neurons to recognize symbols
A group of Russian and foreign computer engineers assembled a primitive "brain" of one hundred artificial neurons based on memristors and taught it to distinguish letters and recognize images, according to the article by Prezioso et al. Training and operation of an integrated neuromorphic network based on metal-oxide memristors, published in the journal Nature (full text is available on the website arxiv.org – VM).
Photo: Sonia Fernandez
"If you try to emulate the work of the brain on a classic computer, you will always face an insurmountable limit in the form of their inefficient architecture. Our memristor technology lacks this drawback, as it is based on a completely different principle of computing, similar to how the brain does it," explains Mirko Prezioso from the University of California at Santa Barbara (in a press release UC Santa Barbara An Important Step in Artificial Intelligence – VM).
Precioso and several other engineers under the leadership of Dmitry Strukov are the creators of a memristor – a special nano-device that combines the properties of a memory cell and a resistor. The memristor "remembers" in which direction and with what force the current flowed through it, which allows it to be used as an analog memory cell in which information is stored in approximately the same way as in nerve cells.
On the basis of these memristors, the authors of the article and other scientists have created several models of nerve cells in the past years, primitive chains of which scientists used to create the simplest computing devices.
In their new work, engineers from the University of California, made, as Strukov himself puts it, a small but very important step – they managed to combine one hundred artificial memristor nerve cells into a kind of "brain on a chip", which they trained to recognize the letters z, v and n in 3-by-3 pixel pictures.
This success is important for several reasons at once. So, the creation of even such a primitive brain from artificial neurons suggests that memristors can be easily scaled. According to Strukov and his colleagues, nothing prevents increasing the number of nerve cells from 100 to 100 billion, and the memristor brain will take up less space than its human counterpart, due to the denser "packing" of neurons, and will work 500 times faster.
The importance is also added by the fact that this design is not a digital, but an analog computer that can solve a whole class of previously unaffordable or extremely time-consuming computational tasks, including the creation of artificial intelligence.
As the engineers themselves admit, memristor analog computers are not a substitute for classical computing devices – now Strukov and his colleagues are working on creating systems that would allow connecting a memristor "brain" to a conventional computer as a coprocessor and exchanging information with it, as well as systems that allow combining several such analog computers.
Portal "Eternal youth" http://vechnayamolodost.ru
14.05.2015