IcaNet1.hpp Source File

00001 #ifndef __IcaNet1__
00002 #define __IcaNet1__
00003 
00028 #include "RealVector.hpp"
00029 
00030 class IcaNet1 
00031 {
00032 private:
00033 
00034 bool                    combined;
00035 bool                    superGaussian;
00036 
00037 double          learnRate;
00038 
00039 RealVector*             w;
00040 double          wNorm2;
00041 RealVector*             dw;
00042 
00043 protected:
00044 
00045 // gMin may approximate sigmoids
00046 virtual double gMin(double u){ return (u - u*u*u); }            // a & b assumed = 1.0 
00047 virtual double gPlus(double u)
00048         {
00049         wNorm2= w->vT_v();
00050         double wNorm4= wNorm2*wNorm2;
00051         return (-u*wNorm4 + u*u*u); 
00052         }
00053 
00054 virtual void learn(RealVector* sample);
00055 
00056 public:
00057 
00058 // default is sub-gaussian
00059 IcaNet1(double rate, bool superType)
00060         {learnRate=rate; superGaussian= false; w=0; dw=0; combined= false;}
00061 IcaNet1(double rate, bool superType, bool combineFlag)
00062         {learnRate=rate; superGaussian= false; w=0; dw=0; combined= combineFlag; }
00063 virtual ~IcaNet1()
00064  {
00065  if( w != 0 )
00066         { 
00067         delete w;
00068         delete dw;
00069         }
00070  }
00071 
00072 virtual RealVector* getW(){return w;}
00073 
00074 // learning
00075 virtual void init(int size);
00076 virtual void learning(RealVector* sample, int nmax=1);
00077 
00078 // out
00079 virtual double forward(RealVector* src);                // maximise abslote value of kurtosis
00080 
00081 virtual void output()
00082         {
00083         printf("Single ICA Net");
00084         }                               
00085 };
00086 
00087 #endif