Non Von Neumann computers are an interesting topic, understanding their usage in the spiking or artificial neural network (machine learning in general). Memristor is a viable building block for implementing analog computers and of course, neural networks in the hardware level. Two main problems (as I can count) for building an architecture based on memristors are their dynamic range and the difference in writing and clearing the memristor (in the terms of linearity as the necessary steps for writing an amplitude to a memristor are not necessarily equal to the ones for bringing it back to its base state).
As we are talking about processing analog data, we deal with the current's amplitude (or maybe phase). But what if we use the frequency instead of amplitude? Instead of multiplying or adding the currents together, we manipulate the frequency in each node. I know that i'm adding (maybe a lot of) complexity to the architecture, but in the meantime, we will achieve (probably) a great dynamic range and linearity in our design. I've found something that is near to my idea in the paper: "STFNets: Learning Sensing Signals from the Time-Frequency Perspective with Short-Time Fourier Neural Networks".
I've some non elaborated ideas in mind to implement this architecture (like using blocks as simple as an analog mixers based on memristors (which maybe some say is backing to our main problems, but I think not necessarily)) and of course some questions remain, like power efficiency or frequency interference.
Anyway, I will appreciate if anyone can participate in this discussion to find a better and more reliable answer to this question.
Thanks.
I think you excited a basic and important question about how can one improve the analog computer to acquire some properties of the digital computer which is the dynamic range. In the analog computers the smallest value is limited by the generated noise in its processor. The other limit is the maximum value which is limited by the saturation or the cut off phenomena. This value is limited by the power supply.
There is the logarithmic function which can be utilized to increases the dynamic range. This is already used in the analog signal processing.
You want to use the frequency of sinusoidal waves instead of the amplitude of the signal to represent the function. Yes the dynamic range of the frequencies is much larger than the amplitude dynamic range.
However there will be two limitations:
The bandwidth of the processing devices
The the resolution in detecting the frequency.
So, the dynamic range may be the ratio of the bandwidth to frequency detection resolution.
This is from the conceptual point of view.
There are also the questions How to perform the mathematical and logical operation in the frequency domain.
I studied the analog versus digital and cam a conclusion that one of the main shortcomings in the analog computers is the absence of the an analog electronic memory similar to the digital memory in the digital computers.
If such memory exists the analog computer will be very attractive in solving many complicated problems.
I would like that you refer to my questions on the RG at the link: https://www.researchgate.net/post/What_would_be_the_impact_of_an_existence_of_an_electronic_analog_memory_on_electronic_systems
Best wishes
I think you excited a basic and important question about how can one improve the analog computer to acquire some properties of the digital computer which is the dynamic range. In the analog computers the smallest value is limited by the generated noise in its processor. The other limit is the maximum value which is limited by the saturation or the cut off phenomena. This value is limited by the power supply.
There is the logarithmic function which can be utilized to increases the dynamic range. This is already used in the analog signal processing.
You want to use the frequency of sinusoidal waves instead of the amplitude of the signal to represent the function. Yes the dynamic range of the frequencies is much larger than the amplitude dynamic range.
However there will be two limitations:
The bandwidth of the processing devices
The the resolution in detecting the frequency.
So, the dynamic range may be the ratio of the bandwidth to frequency detection resolution.
This is from the conceptual point of view.
There are also the questions How to perform the mathematical and logical operation in the frequency domain.
I studied the analog versus digital and cam a conclusion that one of the main shortcomings in the analog computers is the absence of the an analog electronic memory similar to the digital memory in the digital computers.
If such memory exists the analog computer will be very attractive in solving many complicated problems.
I would like that you refer to my questions on the RG at the link: https://www.researchgate.net/post/What_would_be_the_impact_of_an_existence_of_an_electronic_analog_memory_on_electronic_systems
Best wishes
Thanks for your comprehensive answer. I've some solutions and debates in mind. Of course, the most chased answer to the analog memories these days is using memristors and phase change memories. But i think we may can find another way and that's what I want to understand more.
In a simple non volatile memory, we trap electrons in its gate. Probably, this mechanism is not very controllable or reliable. Another solution is of course using memristors which as I said has some shortcoming in the terms of linearity and dynamic range.
There is something that I've not seen anywhere else, to design a FET that has connected gate and back-gate memristors simultaneously. Here, we may be able to mimic the feed-forward and backward in a neural network. Maybe an NVM with a back-gated memristor helps us (I've no deep vision here).
Anyway, I think memristors are the best candidates, but improving them is another topic. Also, photonic analog computers should be considered as another ultimate solution.
An example for Analog
computer is the
Brain. If you look
at an EMG signal
you will find that
it is an FM signal.
This is indeed a very interesting idea. On the one hand we could get rid of some of the limitations of amplitude-based coding, but the probelm then is that in order to perform computation you need to constantly cycle up and down (you need an oscillating waveform), which is very roughly equivalent to constantly flicking a digital gate. In general we want to avoid having to flip many gates many times before we get an answer, so is this something with a sufficiently good runway to match purely digital approaches?
This is something we are working on within the framework of a large UK-based project (FORTE - http://www.forte.ac.uk/) although I'm in a different division looking at things like charge-limited computing.
- Badges
- Science topic