Multivariate Gaussian (normal) pdf
defined as
evaluated at feature vector , with mean vector and covariance matrix satisfying:
and
- Also written as or , where symbol means ``has distribution''
-
- Parameters and describe center and spread of data
- Gaussian pdf has several excellent properties:
- ``Safest bet'' amongst all pdfs with given and
- Closed under multiplication (within scale factor):
- Closed under convolution (i.e. sum of independent Gaussian variables is Gaussian):
- Consequently, for linear systems, in out:
if
- BUT: (mixture Gaussian instead)
- Marginal and conditional densities of are
Full covariance Gaussian
- is general matrix model has parameters
- Models data as ellipsoid-shaped cloud ``at an angle'' allows correlations between feature components
- Most powerful Gaussian model, but needs lots of training data to properly estimate parameters, especially if high
Diagonal covariance Gaussian
- is diagonal matrix model has parameters
- Models data as ellipsoid-shaped cloud aligned with axes correlations between feature components ignored
- Good compromise between modelling power and training data hunger, especially when used in mixture models
Spherical covariance gaussian
- (scaled identity) model has parameters
- Models data as spherical cloud
- Simplest Gaussian pdf useful when data is very scarce or computational speed and storage is an issue
转载地址:http://courses.ee.sun.ac.za/Pattern_Recognition_813/lectures/lecture03/node1.html