News for hackers. Stuff that matters.
Re:Power Method for Approximating Eigenvalues [pdf] by xchip on Wednesday December 27, 2017 @10:12AM (#16015133)
What are you all computing eigenvalues for? I am curious.
Re:Power Method for Approximating Eigenvalues [pdf] by noelwelsh on Wednesday December 27, 2017 @10:20AM (#16015194)
PageRank is computing the principal eigenvector of the transition matrix describing the connectivity of the web. PageRank is no longer, AFAIK, the primary signal that Google uses but it's an idea that was worth $Billion at least.
Re:Power Method for Approximating Eigenvalues [pdf] by hyperbovine on Wednesday December 27, 2017 @01:10PM (#16016618)
Re:Power Method for Approximating Eigenvalues [pdf] by acidburnNSA on Wednesday December 27, 2017 @03:54PM (#16017949)
In a nuclear reactor, neutrons are born (from fission or n,2n reactions) and neutrons are lost (by diffusion or absorption). The rates of these processes are all dependent on how many neutrons there are (call it X), and by conservation of neutrons, the mechanisms can be written:

Losses * X = Gains * X

or equivalently: (Losses - Gains) * X = 0

As it turns out, when you discretize this (matrix) equation over some spatial mesh (and sometimes energy and angular meshes as well), that is a classic eigenvalue problem AX=0 or (L - lambdaG)X = 0. You can just write down the loss and gain terms in every spatial mesh point and pass the matrix into an eigenvalue solver like the power method and you'll get a value for lambda and X. X (the eigenvector) is your neutron distribution in the reactor, it tells you where the neutrons are, where they're going, and how fast they are. This determines how much fission power is being produced where and how much each material is transmuting into fission products and whatnot. The eigenvalue lambda is a number that tells you how critical or not the chain reaction is. If it's 1.0, you're critical. If it's less than 1 you're exponentially growing (going up in power or exploding), and if you're above one, you're exponentially decaying.

And that my friends is how computing eigenvalues is the key to nuclear engineering.

Re:Power Method for Approximating Eigenvalues [pdf] by phunge on Wednesday December 27, 2017 @10:28AM (#16015258)
They're also used in spectral clustering, which is a powerful method for clustering data (see Ng 2001 for some pictures)
Re:Power Method for Approximating Eigenvalues [pdf] by efavdb on Wednesday December 27, 2017 @10:18AM (#16015175)
One example: A common ml algorithm called principal component analysis makes use of the eigenvectors and eigenvalues of a data set's correlation matrix. The dominant eigenvectors can be used as a basis for data compression and the size of their eigenvalues determines the accuracy of the compression.
Re:Power Method for Approximating Eigenvalues [pdf] by hackcrafter on Wednesday December 27, 2017 @04:30PM (#16018192)
With high dimensionality data (think many columns), plotting the first two eigenvectors provides a fantastic clustering of the rows.

In genetics for example, this plot may separate samples with different ethnicities (see PCA on SNPs).

Re:Power Method for Approximating Eigenvalues [pdf] by antognini on Wednesday December 27, 2017 @03:08PM (#16017610)
The eigenvalues of the Hessian of a neural network describe the curvature of the loss function. It is generally believed that in NNs trained with stochastic gradient descent regions with high curvature tend to generalize poorly.