Neural network discussion

[Flare09]

Hey everyone, I had an idea for a hardware neural network. Upon searching the site for memristor there were no results. I wanted to run the idea by someone more knowledgeable to see if the circuit would work.

The basic premise is a modular circuit consisting of a single memristor output connected to ten or so transistors base pins, Let's go with N type. When a signal (from a sensor, let's say photodiode) hits the memristor, the output should trigger one or two transistors. Each emitter would be connected to another memristor. This would then cascade into a specific path to an endpoint.

The modularity aspect comes into play where you can connect another single memristor/transistor network to each emitter, growing the network.

What do you think about this hardware based nerve cell?

 

[VEBill]

I can recall that Memristor-based simple Neural Networks was an active topic decades ago.

Just now: When I typed Memristor into Google, it offered ...Neural Network as an autocomplete. Plenty of hits.

 

[Flare09]

I did find a few hits on the custom Google search available by the site, but i did not in the site search. Either way, does anyone think the idea would work how I believe it would?

 

[IRstuff]

I think the memristor is a solution looking for a problem. Almost 20 years ago, when I was first introduced to ANNs, there were patents issued on similar sorts of devices. Currently, I think a memristor device would be quaint, but no one wants a device they can't interrogate or duplicate. Certainly, while a single node might in interesting at the junior high school level, wiring together dozens of these things would be tedious and extremely space inefficient. You'd need power supplies, oscilloscopes, and logic analyzers just to see what's going on.

Moreover, the weighting device isn't really the bottleneck, it's the algorithms wrapped around it. Google's TensorFlow, among others, are infinitely easier to experiment with, and can provide intermediate results and data, and are extremely repeatable and controllable.

Furthermore, people built multi-element memristor arrays over 4 years ago; even if this idea had merit, it's absurdly late into the market.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert!

https://www.youtube.com/watch?v=BKorP55Aqvg

faq731-376 forum1529 Entire Forum list

http://www.eng-tips.com/forumlist.cfm

 

[Flare09]

The one thing about the memristor I found appealing for using it in this application was that small defects would change the properties of some of them. These imperfections are desirable to somewhat emulate biological counterparts.

I have come across the 6/8/12 dip packages for the arrays. But I think the design of individual packages I am imagining can be manufactured as array on the die. Considering transistor technology is down to 7nm, a 5 inch wafer could house enough to easily emulate a small insect or that fully mapped worm.for small scale experiments, individual networks could be produced in dip or smt.

But you do have a point that I considered, how to measure output and inputs. There needs to be a type of pre-procesessor to sort and filter information, akin to the prefrontal cortex. As well as a post processor to translate specific output paths into desired outputs like movement triggers

 

[IRstuff]

While feature sizes have decreased, bonding pads have not. Moreover, a 7-nm feature sized chip costs at least $500k in NRE, not counting the design itself.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert!

https://www.youtube.com/watch?v=BKorP55Aqvg

faq731-376 forum1529 Entire Forum list

http://www.eng-tips.com/forumlist.cfm