sans-hubris: What's this neural network thing you're working on?

[jemfinch]

I noticed in the General forum that you were working on a neural network system. What's it for?

Jeremy

[sans-hubris]

This was a long time ago, but yeah, I wrote a neural network for a class I had. Unfortunately, the hard drive I had all my code on died.

Neural networks aren't that hard to write though. I could write another one really quickly if I wanted to.

Hehe, it might be helpful if I explained what a neural network is.

A neural network is an artificial intelligence concept. It's a way of modeling nervous systems (including the brain) in software. Basically, you construct software neurons (i.e. they, for the most part, behave like real neurons, except in software) because neurons are incredibly simple, and so incredibly easy to mimic in software. Once you've constructed them, you have to connect them together and get them to communicate with each other. That communication is all mathematical, but no one's been able to prove why it all works.

The beauty of neural networks is that they are fairly flexible. They actually learn, but they're very slow at it (just like us.) After I wrote my neural network, I tried teaching it to solve the XOR problem, and it took forever (not literally, but I stopped counting after a few days.) Although, I could teach it the alphabet pretty quickly, and it did really well on that.

Now, there are a number of different types of neural networks, but the two most common ones are feed forward neural networks (FFNN) and self-organizing maps (SOMs.)

FFNNs are layers of connected neurons, with an input layer, an output layer, and an arbitrary amount of middle layers (although, you usually only need one.) You train FFNNs for very specific needs. Once trained, you can, essentially, give them a question within the field you trained it in, and it'll try to give you an answer, along with its assurance of the correctness of that answer.

SOMs are layers of neurons that loop back on themselves. These neural networks are used to organize data sets. You throw the data into the SOM at random neurons, and it will put all similar data close to each other.

[bwkaz]

So an FFNN would be good at, for example, voice recognition type stuff, right?

Of course, this is just based on my experience with one specific voice recognition chip, so it could be wrong, which is why I'm asking. ;)

This chip (the HM2007) had a mic input, and when it was polled, it'd record for like 2 seconds or so (the time was configurable), then compare the recorded sound against its "database" of pre-recorded sounds. Then it'd output which item it thought it was (the items were generally prerecorded words, but it would work for pretty much anything), and the certainty that the answer it gave was right. Which sounds like an FFNN application to me...

[Strike]

Originally posted by bwkaz
So an FFNN would be good at, for example, voice recognition type stuff, right?

Of course, this is just based on my experience with one specific voice recognition chip, so it could be wrong, which is why I'm asking. ;)

This chip (the HM2007) had a mic input, and when it was polled, it'd record for like 2 seconds or so (the time was configurable), then compare the recorded sound against its "database" of pre-recorded sounds. Then it'd output which item it thought it was (the items were generally prerecorded words, but it would work for pretty much anything), and the certainty that the answer it gave was right. Which sounds like an FFNN application to me...
Yes, it's quite good at it.

I made one :) (in MATLAB, which is kinda cheating, but whatever :))

Actually, that's not true, I made a noise cancellation system. But anything that wasn't part of the voice was considered "noise", so it's basically a voice recognition system. Perhaps not in the sense you meant, though.