GaussClass Gaussian classifier

GaussClass.kr(in, bufnum, gate)

A Gaussian classifier, which classifies an input vector as belonging to one of the gaussian distributions defined in a specially-formatted Buffer.

in - input signal, a multichannel signal specifying a co-ordinate in the space (i.e. a vector).

bufnum - the buffer in which the shapes and weights of the gaussian components are specified.

gate - the classifier is only active when this is greater than 0 (otherwise, previous value is held constant). Its default value is 1.

The Buffer should be single-channel. Its exact format is specified towards the bottom of this file. If you have the MathLib quark installed then you can use the convenience function GaussClass.classesToFloatArray() to create a FloatArray suitable for loading to a Buffer.

The following example creates two-dimensional data with three classes:

(

~classes = [

[ // First class's mean, covariance, weight:

[2, 8],   [[1, 0], [0, 1]],   0.3

],[ // Second class's mean, covariance, weight:

[8, 2],   [[1, 0], [0, 1]],   0.3

],[ // Third class's mean, covariance, weight:

[8, 8],   [[0.75, 0.5], [0.5, 0.75]],   0.4

]

];

~serialised = GaussClass.classesToFloatArray(~classes);

)

// Now let's use it:

s.boot;

b = Buffer.loadCollection(s, ~serialised);

(

x = {

var rate = 20, input, result, gate;

//input = {LFNoise2.kr(rate).range(0, 10)}.dup(2); // Our "query point" wanders around in space

input = [MouseX.kr(0, 10), MouseY.kr(0, 10)]; // Or you can wander yourself using the mouse

gate = Impulse.kr(rate);

result = GaussClass.kr(input, b, gate);

input.poll(gate, "input");

result.poll(gate, "result");

Out.ar(0, SinOsc.ar(result.linexp(0, ~classes.size-1, 440, 880), 0, 0.1).dup); // sonify

}.play(s)

)

x.free;

________________________________

THE FORMAT OF THE BUFFER:

The Buffer should be single-channel and hold data in the following order, once for each class: 

 - N floats: the mean vector; 

 - N*N floats: the inverse of the covariance matrix; and

 - 1 float: the weight of the component divided by the square root of the determinant of the covariance matrix.

N is the dimensionality of the data space. The length of the Buffer is therefore C * (N*N + N + 1). GaussClass.kr will determine the number of classes from the size of the Buffer.