ICA IMPLEMENTATION

• Assume the probability density functions of the underlying independent components are Gaussian or super-Gaussian.

  
  

• Choose a sigmoidal function y=g(u) and get the transformation matrix by maximizing the mutual information of g(u1), g(u2)...

  
  

• Change the training feature vectors into the new space set, then get the mean and covariance matrices for each class.

  
  

• Change the testing feature vectors into the new space sets separately and calculate its distances to each class and give classification.