Hardware Neural Network

If you want to do a HW realization of a neural network your best bet is to do it in Field Programmable Gate Arrays. Try googling FPGA and Neural network:

FPGA Implementations of Neural Networks - a Survey of a Decade of ...

and about 17k more hits.

Have fun, Ed V.

"Daniele" a écrit dans le message de news: snipped-for-privacy@z35g2000cwz.googlegroups.com...

Hi Danielle,

You may have a look at the "Neural Stamp" project I've published in the Circuit Cellar magazine some time ago (January 2000, issue #114) : it provided 8 analog inputs, an hidden layer of 16 neurons, an output layer of

Friendly, Robert Lacoste

It is definitely true that floating point capabilities are not needed for the real-time implementation of a network which has already been trained. They can be an advantage in the training process however.

I developed a real-time multi-layer perceptron implementation around

The training was done on Sun workstations and took many days.

A DSP is exceptionally well suited to the task, because each "neuron" can be implemented as a repeated multiply-accumulate-with-data-move instruction followed by a table lookup for the sigmoid function.

The following publication describes the work:

Real-Time Portable Multi-Layer Perceptron Voice Fundamental-Period Extractor for Hearing Aids and Cochlear Implants. JR Walliker & I Howard. 1989

John Walliker

At some point, someone in the university will ask "how is your work new," particularly if you are a graduate student. So I wonder how you can distinguish this as something beyond the articles by James Albus in the summer 1977 Byte magazine on a CMAC? (That summer, I implemented his algorithm on a TI handheld calculator.)

I half- recall seeing a reference to work on inhibitors in real neurones that said the untreated neurone (from what? can't remember) had a 1-2% chance of triggering with no input, so it would seem 8 bit should be at least similar to real- life. But a 10 or 12 bit lookup shouldn't break the bank with most processors or FPGAs.

Paul Burke

of

th

ily

Thanks for the hints and the link. I already had thought about using an 8-bit table for mapping the sigmoid, so this is a good confirmation.

virgolette -

I have to train the network and set the heights without using a calculator, but writing directly in the memory of the MC/DSP.

Friendly,

Daniele

I know it's an academic darling, but has anyone ever done anything genuinely useful with nn technology?

John

On a sunny day (Thu, 01 Mar 2007 08:03:47 -0800) it happened John Larkin wrote in :

Speach recognition, look up dragon natural speaking. Also look up Liaw and Berger. Now ho is asking cryptic quatrions? Cannot you type neural net in google?

The problem with NNs is that you can't see why they work. Thus although they can provide neat results, you have to verify them afterwards, e.g. speech recognition. You can't prove that they're going to work in any given case without trying it. As Deming said, "You can't test quality into a product."

Cheers,

Phil Hobbs

On a sunny day (Thu, 01 Mar 2007 15:02:14 -0500) it happened Phil Hobbs wrote in :

This is not completely correct. I suggest you look up Berger & Liaw. For now it also is a mathematic question, so you can calculate what comes out. Anyways Berger & Liaw came up with a betetr neuron model. It think this is now used to find snipers and gun types(??), but hard to get data, more likely submarine detection, as it is Navy financed. Better models is what we need.

Hmm. So if you have one of these B&L gizmos (about which opinion seems to be seriously divided), and it's been trained to recognize my speech, how are you going to show that it'll recognize yours without trying it?

Cheers,

Phil Hobbs

The same problem exists for all complex pieces of s/w. There is no general way to prove that they are bug free or will do what they are supposed to do.

Well, no, that's not true. There are unit tests and so forth, and you can trawl through the code and see how it's organized. Try doing that with a neural net. NNs are cool, don't get me wrong, but I _hate_ 3 AM phone calls.

Cheers,

Phil Hobbs

On a sunny day (Thu, 01 Mar 2007 16:59:37 -0500) it happened Phil Hobbs wrote in :

Sing a song into it?

out.

My point.

Cheers,

Phil Hobbs

Oh, there are lots of hits, too many in fact. I was just wondering if any practical products have resulted. The cited applications seem to be stuff like spam detection, language translation, and pattern recognition, processes that really don't expect consistant accuracy.

I'd be reluctant to trust anything serious, like a control system that mattered, to an algorithm whose corner cases are undefined and probably not testable.

I've worked with a few academics that would shout "neural network!" (in one situation, two did it in precise unison) in response to nearly any problem they didn't have an analytic approach to. The suggestion was often beyond absurd.

John

Is there _any_ pattern recognition technology that doesn't suffer from the same problems? At least for complex signals? This is not to say that a truly refined technology (ideally) shouldn't be able to deliver consistent/provable/testable performance, or be open to inspection. But then the finest pattern recognizers - at least, in discriminating the most complex signals in significant noise - (trained people) don't meet this test, either. We have a long way to go in developing a technology that has the fine attributes that you seek.

-f

No, but it seems fundamental to me that this is the only class of problems - ones where a substantial error rate is acceptable - that nn's are suited for.

John