Printable View
Hello everyone, well recently i've become entranced in this book entitled: AI, The Tumoltuous History of Artificial Intelligence by Daniel Crevier. If you haven't read it you should, its a great book, very unique in its approach.
Well anyways, me being the person i am as far as jumping from one thing to another...
Are there any simple approachs to some AI topics, such as neural networks, that i could get started on? Or perhaps another book which covers a more technical approach?
Thanks!
Honestly, I think neural networks are a bad place to start. I'd start with some simple search-based stuff and a background on predicate calculus. Try "Artificial Intelligence: A New Synthesis" by Nils Nilsson.Quote:
Originally Posted by Junior89
Neural networks are really a very powerful nonlinear interpolation technique for high-dimensional functions. They were inspired by biological neural networks, so they were originally considered "AI" but the traditional methods like backprop are pretty much nuts and bolts these days. The cool stuff is happening elsewhere. Nilsson's book does cover them, though, so you can read about them there.
Thanks!
> Neural networks are really a very powerful nonlinear interpolation technique for high-dimensional functions.
Actually the simplest case of a Perceptron is a linear example, but I agree with your post anyway. Start with some simple search algos, decision trees, etc...
Only a single-layer perceptron. A linear, multi-layer perceptron is equivalent to a single-layer perceptron. Therefore there is no point in creating a multi-layer, linear perceptron. So all multi-layer perceptrons are by definition nonlinear.Quote:
Originally Posted by Perspective
Actually, it's the perceptron itself that is linear. When you move to something non-linear (eg. sigmoid) it's no longer a perceptron.
You're right that there's no point in having multiple layers as they are always equivalent to some single layer perceptron though, but that's a different point.
Perceptrons... haha, i read about them too i think. =P
But i'll start small ;) Like search algorithms, decision trees, etc.
And work my way up =)
Despite that, the term "multi-layer perceptron" continues to be used for multi-layer NONLINEAR networks. See this:Quote:
Originally Posted by Perspective
http://en.wikipedia.org/wiki/Artific...yer_perceptron
This class of networks consists of multiple layers of computational units, usually interconnected in a feed-forward way. Each neuron in one layer has directed connections to the neurons of the subsequent layer. In many applications the units of these networks apply a sigmoid function as an activation function.
Interesting, I've never heard the term perceptron as meaning something non-linear. Though wikipedia does contradict itself with the definition of a perceptron:
http://en.wikipedia.org/wiki/Perceptron
The perceptron is a type of artificial neural network invented in 1957 at the Cornell Aeronautical Laboratory by Frank Rosenblatt. It can be seen as the simplest kind of feedforward neural network: a linear classifier.
The whole motivation behind multi-layer non-linear networks is that perceptrons can only solve (classify) linearly seperable problems.
But anyway, it's just a terminology issue.
I don't think it's a contradiction, I think it is bad terminology which Wikipedia didn't invent. A "multi-layer perceptron" really isn't a derived class of "perceptron," it's a whole other thing. Somebody probably misspoke or used bad terminology at some point and the nomenclature stuck.Quote:
Originally Posted by Perspective