Unraveling the Mystery of "Neural Networks"

[Brian_C]

Can anyone explain to me what this "neural network" nonsense is all about? Everyone and their mother is doing research in this area, but nobody seems to know what "neural networks" are or why they would be useful. It sounds like pure hype to me.

 

[story645]

Can anyone explain to me what this "neural network" nonsense is all about?

c&p from my psychology paper:
In its most basic form, a neural net is a a set of units connected by statistically weighted links. (Russel & Norvig, 1995, p. 567). The units are mathematical stand ins for neurons-they do all the actual processing based on some set of inputs (including a current activation level) and they give back some output (including a new activation level). The weights on the links determine how important that specific link (signal) will be in the neurons computation, and these signals are then passed on to the next unit(s) in the chain, continuing until the signal has passed through whatever hierarchy it was supposed to. (Russel & Norvig, 1995, chap. 19) The weights are adjusted through a series of trials so that each node can learn which signals it needs to pay attention to and which ones it should filter out, making neural networks very suited to studying attention tasks, wherein every neuron has to decide what computational weight it wants to attach to each activation potential traveling through it.

Basically, people like neural nets 'cause they're good for studying/modeling networks with feedback.

 

[Proton Soup]

Basically, people like neural nets 'cause they're good for studying/modeling networks with feedback.

you can also train them to do things without actually understanding how they do it.

i wonder if Toyota uses them to control any of their automotive systems ?

 

[zoobyshoe]

c&p from my psychology paper:
In its most basic form, a neural net is a a set of units connected by statistically weighted links. (Russel & Norvig, 1995, p. 567). The units are mathematical stand ins for neurons-they do all the actual processing based on some set of inputs (including a current activation level) and they give back some output (including a new activation level). The weights on the links determine how important that specific link (signal) will be in the neurons computation, and these signals are then passed on to the next unit(s) in the chain, continuing until the signal has passed through whatever hierarchy it was supposed to. (Russel & Norvig, 1995, chap. 19) The weights are adjusted through a series of trials so that each node can learn which signals it needs to pay attention to and which ones it should filter out, making neural networks very suited to studying attention tasks, wherein every neuron has to decide what computational weight it wants to attach to each activation potential traveling through it.

Basically, people like neural nets 'cause they're good for studying/modeling networks with feedback.

A guy I met here in San Diego described them as computer programs that can "learn" in some way, shape or form. He, himself, was teaching a neural net to play poker, or, perhaps, allowing it to learn how to play poker would be a better description because all he was doing was waiting while the program ran for weeks by itself.

 

[story645]

A guy I met here in San Diego described them as computer programs that can "learn" in some way, shape or form.

That's because they're feedback functions, so outcomes are used to reweight paths until the weights are settled such that a "correct" output is reached.