Connect with us


Engineers are able to make chips in smartphones 1000 times more powerful




A team of engineers from University College London (UCL) has developed a new approach to building neural networks based on memristors that work virtually error-free. Until now, it was believed that the use of memristors and the accuracy of calculations in the construction of neural networks are incompatible.

The transition of artificial intelligence systems from transistor hardware to memristors will increase the energy efficiency of AI by 1000 times, and this will lead to the rapid emergence of powerful neuromorphic chips almost everywhere – from smartphones to industrial systems, according to Tech Xplore.

The UCL team figured out how to solve this problem and the simulation confirmed that it was correct. The decision turned out to be surprisingly simple. Scientists forced memristors to work in several subgroups of neural networks and averaged their calculations. Thus, the overall performance decreased slightly, but the number of errors decreased to almost zero.

In addition, scientists have tested the approach on several types of memristors and found that accuracy increases with the use of any model, regardless of material or manufacturing technology.

An open method of combating errors can be the basis for the development of a new generation of artificial intelligence.

The emergence of memristic neural networks or neuromorphic chips with energy efficiency 1000 and more times higher than in current transistor systems, will effectively train neural networks in general without connecting to external resources. Their internal resources will be enough for this. Obviously, this opportunity will turn more than one industry.

And this resource is provided by the very nature of memristors – they are also called “memory resistors”, because they remember the amount of electric charge that flowed through them even after shutdown. Thus memristors work not only in the binary code consisting of zeros and ones, but also on several levels from zero to one at the same time.

This means that each bit can hold more information. And given the fact that the operational data is processed and stored in one place, they do not need to be constantly sent to memory and receive from it during calculations, all this increases the efficiency of such systems by orders of magnitude in comparison with transistors.

The authors of the project claim that at this stage their AI is equal to existing neural networks and performs tasks at the same level, but this is only the beginning of promising development. Scientists promise to build the first functioning model based on memristors in three years.


Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Popular on Gadget Tendency