The development of a civilization often takes place by cycles. When one is closed another one opens and sometimes it has been closed before.
What I mean? We all know that our civilization has been based on digital developments for years now. The digital race has led to abandoning technologies that seemed to have no future but which, it seems, at least, could soon be back in vogue.
According to a recent IBM study (published in Nature in 2018) the use of "analog memories" could be a solution to the efficiency needs of the neural networks used for Artificial Intelligence.
Researchers at the IBM Research AI team were able to demonstrate that it is possible to use peripherals with analogue memory (which use continuous electrical signals instead of the more well-known binary signals composed of 0 and 1) for the deep Learning achieving the same accuracy that is achieved using digital graphics processors.
One might ask, given that the accuracy achieved is the same, what is the novelty, why should analogue memory be favored. The answer is to be found in the way digital processors work, in their architecture and in the need to move large amounts of data to carry out the necessary training of neural networks, all these factors must be taken into consideration.
The development of Artificial Intelligence is in fact based on the enormous and ever-increasing amount of data that must be collected, analyzed and processed. To do this, especially during the deep Learning, the data is moved between memories and processors and this means use of time and energy. The study of new architectures (often referring to natural elements, such as the human brain) has made it possible to improve the process of deep Learning, "moving" some parts of the memory to the data, simplifying and speeding up data transfer operations. Now, according to IBM researchers it is possible to use analogue memories that allow savings in terms of dissipated energy.
One of the problems that had prompted researchers to use digital memories was the lack of accuracy of analog systems, which however seems to be at least partially overcome. According to the most recent studies, in the next ten years the use of analog technologies in the AI will allow to reach an efficiency one thousand times higher than the current one.
As we have known for a long time, the human brain is an almost perfect machine, also from the point of view of energy consumption and heat dissipation. If you look at the way the human brain works and try to say if it is an analogue or digital "processor", you would find that it is not one another but uses processes similar to one or all another depending on convenience and function. Once again we see how the study of the brain helps us understand how to improve the processors.
To learn more:
- https://www.ibm.com/blogs/research/2019/02/ai-hardware-center/
- https://www.ibm.com/blogs/research/2018/06/future-ai-better-compute/
- https://www.nature.com/articles/s41586-018-0180-5
- https://www.ibm.com/blogs/research/2018/06/approximate-computing-ai-acce...
- https://www.quora.com/Is-the-human-brain-analog-or-digital
