nphrk

It is the distribution of the average of n binomial distributions (taking on values x±1)†, which according to the CLT converges to a Gaussian as n→∞.

[†] x is where it is centered.

The copyright issue can be a bit tricky. Assuming that Instagram goes on and sells users' pictures - who's liable if somebody uploads a picture which he/she doesn't own, and then Instagram goes on and sells that to a third party?

Great! I love the service, I miss only an app for my Android tablet. Keep up the good work.

The article uses the magnitude of the coefficients, which is computed using both the real and the imaginary part.

This is a nice way to see how the DFT is computed, however I find the view of the FT as a change of basis as even more important - generalizes easily to other bases and and one can understand easily wavelets and their advantages. Basically, the sinusoids form a basis of the vector space of functions (every 'non-pathological' function can be written as a possibly infinite sum of them) and the numbers computed by the FT are coefficients for the respective basis vectors - the magnitude of these coefficients is interpreted as the strength of the corresponding wave in the original signal.

Another way to see the FT is as the basis where the convolution operators are diagonal - this is used in image processing, where computing the FFT of a filter + entry-wise multiplication can be much faster than running the convolution at each pixel of the input image.

Could you recommend some introductory books? I've always felt bad not having clue about art.

Just landed!

Very interesting, I have only one nitpick (maybe for the paranoid only) : why not show the cumulative distribution but pick 200$ as a threshold?

Well not quite. While SVMs gained a lot of popularity for having nice properties e.g.

1) a convex problem which means a unique solution and a lot of already existing technology can be used

2) the "kernel trick" which enables us to learn in complicated spaces without computing the transformations

3) can be trained online, which makes them great for huge datasets (here the point 2) might not apply - but there exist ways - if someone's interested I can point out some papers)

There is an ongoing craze about deep belief networks developed by Hinton et al. (who is teaching this course) who came up with an algorithm that can train them reasonably well (there exist local optima and such, so it's far from ideal). Some of the reasons they're popular

1) they seem to be winning algorithm for many competitions / datasets, ranging from classification in computer vision to speech recognition and if I'm not mistaken even parsing. They are for example used in the newer Androids.

2) DBNs can be used in an unsupervised mode to _automatically_ learn different representations (features) of the data, which can be then used in subsequent stages of the classification pipeline. This makes them very interesting because while labelled data might be hard to get by, we have a lot of unlabelled datasets thanks to the Internet. As what they can do - see the work by Andrew Ng when they automatically learned a cat detector.

3) DBS are "similar" to biological neural networks, so one might think they have the necessary richness for many interesting AI applications.

Well not quite. While SVMs gained a lot of popularity for having nice properties e.g.

  1) a convex problem which means a unique solution and a lot of already existing technology can be used
  2) the "kernel trick" which enables us to learn in complicated spaces without computing the transformations
  3) can be trained online, which makes them great for huge datasets (here the point 2) might not apply - but there exist ways - if someone's interested I can point out some papers)

There is an ongoing craze about deep belief networks developed by Hinton (who is teaching this course) who came up with an algorithm that can train them (there exist local optima and such, so it's far from ideal). Some of the reasons they're popular

  1) They seem to be winning algorithm for many competitions / datasets, ranging from classification in computer vision to speech recognition and if I'm not mistaken even parsing. They are for example used in the newer Androids.
  2) They can be used in an unsupervised mode to _automatically_ learn different representations (features) of the data, which can be then used in subsequent stages of the classification pipeline. This makes them very interesting because while labelled data might be hard to get by, we have a lot of unlaballed datasets thanks to the Internet. As what they can do - see the work by Andrew Ng when they automatically learned a cat detector.
 3) They're "similar" to biological neural networks, so one might think they have the necessary richness for many interesting AI applications.