Thursday, March 5, 2015

TOWARDS THE LOW POWER SYSTEM FOR NEURAL NETWORKS

Requirement of low power system in the field of neural network prompts the evolution of new technology to tackle with the problem of significant power consumption.

When large number of neurons in a circuit deals with the particular application then power consumption is high enough to scath the circuit. So a low energy consumption circuit for neural networks has been developed by using ferroelectric memristor. First ferroelectric memristor was developed by Panasonic.

Beauty of the devices using this technique is to treat digital signal as analog data. This analog data is stored continuously as resistance on a CMOS circuit. As analog data is processed directly over here so it is obvious that power consumption diminute drastically.

Generally data storing templates always exist in neural networks so there is requirement of some kind of memory element, this need is undeniably accomplish by ferroelectric memristor.

Continuous change in the value of resistance of ferroelectric memristor occurs by accordingly change in the applied voltage. It behaves as a memory device when there is no voltage applying which leads a maintained resistance. And this fixed resistance acts as stored templates. As analog signal is processed in this technology so any intermediate value can be stored as data rather than binary value 1/0 due to this fact it is capable to store more information than the digital memory.

According to the large amount of information stored in memristor it is possible to lessen the power consumption as the dealing with analog signal proven its vitality in up surging the amount of data to be processed. This innovative development leads to an  efficient neural network in terms of accuracy and power consumption.