What are some good books/lectures on artificial intelligence?

Superintelligence by Nick Bostrom

>>8504436


What computers still can't do by Dreyfuss.

File: 0136042597.jpg (84KB, 500x648px) Image search: [Google]

84KB, 500x648px

I don't wanna be that guy, but I'm gonna be that guy....

>>8504452
This book is great. One of the only textbooks that pulls off humor well

>>8504450
>>8504451
fags

>>8504452
Yeah, this is pretty good. The latest edition is the 3rd, which is getting old by now. But the only major changes in the field that I'm aware of since the 3rd edition was written were that connectionist approaches (neural networks) are taken a lot more seriously. Goodfellow and Bengio are big names in that area, and they wrote a free book: http://www.deeplearningbook.org/

>Are neural networks a good start?
Unless you have a particular problem in mind, then it's as good a place to start as any.

Andrew Ng's coursera course on machine learning is recommended widely because it's very good. If you want to be introduced to neural nets specifically, it's important that you understand them in the broader context of machine learning - it'll take less effort to learn and you'll understand what's appropriate for what task. So this course is good for that. And if you only want something on artificial intelligence in general, this course will introduce you to the serious math and optimization behind most of what we do.

>>8504436
sauce

>>8504436
I want to paizurai a tree

>>8504436

Neural networks are nice.
Neural networks are used for interpolation problems. Consider an N by M black and white bitmap.
This forms a space [math]\mathbb{I}^{N M} [/math] (where [math]\mathbb{I} [/math] is the range of pixel values.

What you then want to do is approximate the subspace of images which have some property, which you do by a fitting a neural net to given points from this subspace. Now you use your approximation to classify images depending on their distance from this approximation.

The above should be enough to give you a starting intuition about simple neural networks.

>>8504739
To continue on this, each image is a vector, and a neuron works by applying a linear transformation on this vector, taking a norm of the result and applying a transfer function.

The transfer function is chosen when we choose our model, likewise with the norm (though the norm is almost always [math] \ell_1 [/math]).

>>8505035
So therefor the only thing to do is optimize the linear transformation.

File: ae5aa46c39bde4dd2de0efdc8a8af3f2.jpg (42KB, 500x448px) Image search: [Google]

42KB, 500x448px

>>8504739
>>8505035
>>8505059
>mfw taken some hardass algebra courses and understood most of it
thanks a lot, I have a great picture to start with

>>8505145
Glad to help, if you want to understand better then take some points in (0-1)x(0-1) and divide them into two classes in a way that a straight line can partition these classes. Find a linear transformation that partitions these classes. This is basically the simplest binary classification problem.

Now try it when the points can not be separated according to their classes by a straight line. For example try the classes (x,y) in A if [math] x^2+y^2 < 1 [/math] and (x,y) in B otherwise.

You will not be able to make the above work with only one neuron (though you should try to see why!)

Now choose at least three tangent lines to the unit circle in the above example. Each of these will partition the points and have all p in A one the correct side, but fail for some B. So non of these partitions are good enough.

Now add one neuron n that takes the outputs of the other neurons as input, choose the linear transformation for n such that it classifies a point p as belonging to A iff all the previous neurons think it should belong to A.

Make sure you actually try this instead of just reading what I wrote and you should be golden. You should be able to do these examples with just some linear algebra and choosing any transfer function (try the step function). Do the above with pen and paper.

Now you know how a neural net does binary classification, what you need to do now is understand how to get your linear transformations algorithmically instead of by guessing.

Take the above example (A-B partitioning) and initialize the neurons somewhat randomly, now look up backpropagation and do it by hand for a number of points. You can choose your points smartly just by looking, see what happens if all neurons agree and are correct, all neurons agree and are incorrect, and the case where not all neurons agree. See how much the system changes and how.

If you have any questions just ask away as long as this thread lives.

>>8504739
>>8505035
>>8506114
Thanks for this, anon. Would you recommend a book(s) for learning and getting good in this?