The concept of dimensions pops up within a multitude of contexts and even in every day life on a regular basis; a line is one-dimensional, a cube has 3 dimensions etc. It may generally be defined as the minimum number of coordinates required to uniquely describe a point within a space. However, different contexts require different definitions. In terms of vector spaces, one valid definition of dimension is the cardinality of the generating set of that space.
To reduce dimensions of high-dimensional data points we need to project them to a not-so-high-dimensional space. To do so one must find components/vectors which maximize variance. These components are called eigenvectors and are the result of a principal component analysis (PCA). High-dimensional data points (e.g. an image or a song) of some dimension are projected to a hyperplane of dimension . This hyperplane is spanned by eigenvectors. It would be nice to make this concept a little more tangible, to visualize it in some way. And, while we are at it, there is no need to restrict ourselves to the visual modality, we can also try and make a high dimensional space (or it’s reduced components) audible as well.
Eigenfaces are probably the most popular eigenvectors.
The face space
A picture is most naturally represented as a matrix of pixel values. However, the rows of the matrix might as well be concatenated into one long row vector, i.e. be represented as points in an ‘image space’. For simplicity’s sake let’s narrow it down some more and consider only gray scale images and only images depicting human faces. Despite these constraints, your ‘face space’ is still huge, for 250x250 pixel images its dimension is just that, 62500.
Eigenfaces are the eigenvectors of the covariance matrix of the face vector space.
Step by step:
Let be a matrix build from row vectors
where the row vector is
The scatter matrix then is
So is a matrix, where is the number of pixels in one picture.
Now the last step is to compute the eigenvectors of , but even if we use small images, will be huge and computation expensive.
No! Fortunately we can make use of the following property:
If is an eigenvector of then:
This means we can compute the eigenvectors of by computing the eigenvectors of the much smaller matrix :
Let be such an eigenvector, then: is an eigenvector of Now normalize and the results are the eigenfaces:
All this can be directly translated to Python (numpy is your friend):
eigen(.) is not specially tailored for faces. As long as the argument is a matrix…
So what happens if our argument is a matrix of sound files? How about some Johann Sebastian Bach?
Well, the problem is:
As written above; the task would be to multiply with , but these matrices are 15G large…
So since Bach was quite productive I chose a very small subset, namely The Well-Tempered Clavier (a reasonable set of 98 .mp3 files). I converted the .mp3s to .wavs using mpg123 and then used the first million frames of every .wav, which then resulted in 1000000/44100 = 22.68 s long eigenclaviers or rather partial-eigenclaviers (slowly eigen* becomes a well-worn prefix).
However I uploaded the first ten. Out of the first ten, these three may be the most interesting: