An artificial neural network is an attempt to simulate the brain. It is made up of connected neurons that have cell bodies (that has a threshold that must be overcome for it to fire), dendrites (each of which accept separate inputs and also have a weight by which the input is mulitplied), and an axon (which can output the same voltage to multiple neurons). (If I have any of this wrong, correct me.) Basically, the brain has to be (if we have no soul) an I/O machine (though a quite complex one): you feed it input, it produces output. The only thing that makes it seem random is that there is so much input, so many variables to be considered that there are countless outputs. The set of final outcomes would be the will. THEORETICAL: If we had a soul, then it would be the will and the brain would be a mere processor/learner directed by the soul. THEORETICAL: If we had NO senses whatsoever, I don''t believe (with my current paradigm) that we could think. There would be nothing to process, and we would be empty dummies, like a network of pipes unconnected to a pump. Otherwise, where would the initial electronic impulses come from. I wish to write an ANN, but I dont want to write it conventionally. The only algorithms I have seen are where the the program already knows the desired outcome and resets the network''s weights so that the desired outcome is produced in later processing. I want to make an AI that can learn on its own. Is there anything in the human brain that causes the neuron''s synapse weights to be reset? I mean...unless we have a soul, there is not a goal from which to set the brain. Well, does anyone have any insight into this? And dont give me a whole bunch of calculus...I am good at math, but I have no calculus experience yet, and I dont quite understand sigma notation yet...

First off, I must say that if you are not comfortable with calculus (particularly partial derivatives) you probably should bone up on it before trying to actually understand what an ANN is doing. With that said, it is possible to write a decent ANN without the calc background (though you will probably have a bitch of a time since you will not know why it does what it does, ect).

If you really want to try an ANN with only input try a kohonen self organizing map -> they are about as close to what you want as you will get.

- mongrelprogrammer

The human brain is a little more complex then just a pure I/O machine. For instance, we can create out own conditions in the brain. This is what creates abstract thought and sets us apart from animals. Also, the neural net takes out the complex synaptic chemistry, which is where most of the action in the brain takes place, so it doesn''t really model a human brain that well.

Boy that was classic... the GameDev.net recent threads listing croped the posting title to...


True AI - Based on the human bra

I thought it was an april fools thing.

But anyways... full ANN are extremly difficult. Not the type of topic that you are going to get on a forum like this. Most research libraries will have much information on the topic. Just spend a few days going through the racks. I have done this approach on several difficult topics and can say that it really helped.

hehe. Bra''s, now there''s a topic...

You don''t need to understand calculus to achieve results with neural nets. You don''t even need to know calculus to know what they are doing. I''m guessing that mongrelprogrammer is referring to his experience with *supervised* learning where knowledge of calculus does come in handy.
However, a lot of people tend to forget about unsupervised learning which, in my experience, is much more flexible anyway and certainly a whole lot more fun to play around with.

Taile: What makes you think ANNs can''t learn on their own? Mine usually do. They have to because most of the time i don''t really know what it is exactly I want them to do. So they have to find their own solutions.

Also, the thing you mention about having to feed an ANN input. You can actually get some really interesting results by training a network, then shutting off its inputs but randomly stimulating synapses. Check this interesting link out:

http://www.imagination-engines.com/aipromise/aipromfs.htm



Stimulate

Now now, fup...

I did remember unsupervised learning as I mentioned kohonen self organizing maps. I still think that knowledge of calc is extremly helpful for all kinds of ANNs, even unsupervised. However, I didn''t go to so far as to say you can''t enjoy them without that knowledge -> I just said that it is helpful (very helpful at times).

Perhaps I should have made myself more clear, it isn''t completly necesary to know calc to do things with ANNs, just helpful : ) I know that when I first started learning ANNs my knowledge of calc made things a ton clearer in regards to how learning rules work and how to speed them up (momentum, deterioration, ect).

- mongrelprogrammer

quote:
First off, I must say that if you are not comfortable with calculus (particularly partial derivatives)

Well, most people use the same sigmoid function, which because of its popularity has been studied into the ground. The function turns out to have a rather trivial answer to a complicated problem and that answer can be had in any AI book worth its weight. Therefore, calc is not truly required, unless of course you plan on using other functions for the sigmoid function, and then the derivatives will get nasty. Very nasty, especially if they dont simplify right :-P

There is another subtle difference between ANNs and the human brain. An ANN has an input stage, a computational stage and an output stage (these are usually separated temporally). The human brain does not. It receives input continuously and provides output continuously and the entire cycle is highly non-linear (as opposed to the linear ANN cycle). Inputs combine with outputs to produce other inputs... internal neural activity is compounded or tempered by further input or by neural activity from other, loosely coupled, portions of the brain. The brain also uses neuro-chemistry to store intermediary results of computation which can be used again later in an entirely different computation for an entirely different affect!

I believe though that what you are asking for is an ANN that can learn it''s weights during operation. This is the problem of sequential, or online, learning. The problem is particularly technical and would be suited to an advanced undergraduate or even postgraduate student. If you want to investigate the problem respond in this forum and I will provide some literature references... they''re rather technical though so consider yourself warned!

Cheers,

Timkin

quote: Original post by Taile
Is there anything in the human brain that causes the neuron''s synapse weights to be reset? I mean...unless we have a soul, there is not a goal from which to set the brain. Well, does anyone have any insight into this?

how about endorphins and all those other neato chemicals? doing/thinking about certain things can cause our bodies to release chemicals to make us feel good, or get anxious/ready to fight, or be relaxed, or become hungry/satisfied... maybe these are what cause the "weights" to be "reset". depending on if the "feel good" chemical is floating around in your brain, your neurons might "tweak" their values and "teach" your brain that this action -> A Good Thing.
i don''t know jack about AI, and only remember a fraction of my biology, but that''s my 2 cents...
timkin: i wonder if that is why that classical conditioning stuff works (you know, pavlov and his doggies)?

--- krez (krezisback@aol.com)

quote: Original post by mongrelprogrammer
Now now, fup...

I did remember unsupervised learning as I mentioned kohonen self organizing maps. I still think that knowledge of calc is extremly helpful for all kinds of ANNs, even unsupervised. However, I didn''t go to so far as to say you can''t enjoy them without that knowledge -> I just said that it is helpful (very helpful at times).

Yes, sorry. I missed that line completely.



Stimulate