Forum websites for AI

[Karamel]

Hi,

I'm Electronics Engineer probably like most of you but I am currently dealing with syntetic data generation with Generative Adverserial Neural networks from wav files database. (https://bhichallenge.med.auth.gr/ICBHI_2017_Challenge)

Actually I don't have any idea which forum websites are using to discuss that kind of problems. Do you suggest anything for me? or Do you interest in with that kind of projects even though you are not a computer engineer? or data scienctist?

Best Regards,
Karamel 

[RJSV]

   I'm not expert, but interested in lower to mid-level solutions.  That is, you can see in the topic list, a thread discussing AI for images, that's Tesla related.  That sets up the problem(s).  Then, they started introducing the hardware, having 12 imbedded CPUs.  Eventually, my level (of interest) the talk getting more direct, a lot of convolution, and even more direct, the (project managers) wanting 'multiply and accumulate answers'.
   That's where I can go to work on problem, in my case doing that with some different approaches.  That could be 'mechanically based'...I've done some speculation on that.  Could be biological,...a bit out of my expertise.
Right now, I'm testing the waters, for using optically interfaced components.
   How can I organize a whole lot of simple optically interfaced components, doing the sub-tasks outlined by the 'big guys'... the ones that simplify my orientation by just wanting to focus on 'convolution' solving.

   Right now, I've started contemplating some sense of a 'liquid suspension' hardware set, to do the mass calculations involved, in AI and image recognition.
Please see also, 'FSD' or 'Fully Self-Driving' topic, that helped me get an overview.

[RJSV]

   So, I know, I'm flailing about, in this explanation (so far).  But here is some more thoughts that might clarify my position:
   Setting aside, for the moment, the bigger picture for neural network, I've followed SOME of the explanations and attempt to distill some of the minutia.  The thing has arrays of pixels, 96 by 96, let's consider that as X,Y and a 'depth' of computation, that provides the 'Z' third axis, for optical processing.
So I've essentially reduced the bigger, more challenging process, (temporarily), and wish to apply the existing framework, often called 'Multiply and Accumulate', by borrowing from that Tesla FSD overview, of the dual IC neural net...How can I translate (downward) into something workable on 'my' (admittedly) simplified level of expertise.
   But, instead of implementing a digital 8 bit by 8 but multiply, and 8 bit accumulator, my discussion veers to the analog realm, using optical 'analog' methods, if possible.  WHEW, that was difficult to state (lol)!
   The obvious 'optical analog' signal is going to be highly linked with DISTANCE, under simple concept: any light emitter having simple reduction in intensity with gate to gate distance, emitting LED, to receiving Solar Cell sensor.  But also, now incorporating simple forms of Pulse Density Modulation, for exploiting the advantages of being a step away from (or free from) the mechanical, or spatial constraints, caused by decay of analog intensity changes (unwanted).
   Now I can go to work, on THAT topographical and layout set of issues.  Setting aside that 96 X 96 array, and just operating on a single (1 pixel) lateral dimension, and a bunch of complexity, going in depth direction. That, to process as 8 bit equiv. in accuracy, doing the classic Multiply - Accumulate.
   MAYBE, in about 32 years, I can  comprehend  what 'convolution' and 'de-convolution' really are!  Meanwhile I'm more comfortable distilling down to that sub-problem...for now.
Others are, quite likely, sticking to the digital solutions.  I'm interested in the opto-electronics route, and learning what are the bounds of performance there...even as I'm aware that likely won't supersede the main body, of research.  That's how hobbies grow!
Thanks. Rick in Hayward

[SiliconWizard]

Generative Adverserial Neural networks? Wow. That sounds like fun.