Generic implementation of D to K dim Self Organizing Map (SOM) More...

#include <SOM.h>

Inheritance diagram for icl::math::SOM:

Classes
struct	Neuron
	SOM internal Neuron struct. More...

Public Member Functions
	SOM (unsigned int dataDim, const std::vector< unsigned int > &dims, const std::vector< utils::Range< float > > &prototypeBounds, float epsilon=0.1, float sigma=1)
	create a new SOM with given data dimension and grid dimensions. More...

	~SOM ()
	Destructor. More...

void	train (const float *input)
	trains the net using given input vector and current parameters epsilon and sigma More...

const Neuron &	getWinner (const float *input) const
	returns a reference of the winner neuron dependent on a given input vector (const) More...

Neuron &	getWinner (const float *input)
	returns a reference of the winner neuron dependent on a given input vector More...

const std::vector< Neuron > &	getNeurons () const
	returns the neuron set (const) More...

std::vector< Neuron > &	getNeurons ()
	returns the neuron set More...

const Neuron &	getNeuron (const std::vector< int > &dims) const
	returns a neuron at given grid location (const) More...

Neuron &	getNeuron (const std::vector< int > &dims)
	returns a neuron at given grid location More...

unsigned int	getDataDim () const
	returns the data dimension More...

unsigned int	getSomDim () const
	returns the SOM dimension (e.g. 2D, 3D ..) More...

const std::vector< unsigned int >	getDimensions () const
	returns the SOM's grid dimension (e.g. 20x40 for a 2D SOM) More...

void	setEpsilon (float epsilon)
	sets the current learning rate epsilon More...

void	setSigma (float sigma)
	sets the current grid distance function standard deviation More...

Protected Attributes
unsigned int	m_uiDataDim
	internal data dimension variable More...

unsigned int	m_uiSomDim
	internal SOM dimension variable ( = m_vecPrototypeBounds.size() = m_vecDimensions.size() ) More...

std::vector< unsigned int >	m_vecDimensions
	internal grid dimensions More...

std::vector< utils::Range< float > >	m_vecPrototypeBounds
	internal bounds for prototype ranges (todo: is it necessary to store them ?) More...

std::vector< Neuron >	m_vecNeurons
	set of neurons More...

std::vector< unsigned int >	m_vecDimOffsets
	internal utility offset vector for each dimension More...

float	m_fEpsilon
	learning rate More...

float	m_fSigma
	standard deviation for the grid distance function More...

Detailed Description

Generic implementation of D to K dim Self Organizing Map (SOM)

Input dimension as well as output dimension is freely configurable, i.e. SOM neurons can be aligned in a K dimensional grid with arbitrary cell count for each dimension. For more convenience an inline wrapper class called SOM2D is also available in "iclSOM2D.h"

Nomenclature

list of neurons: $W=\{n_1,...,n_N\}$
best matching neuron index:
prototype vector of neuron : of dim
grid position vector of neuron : of dim
learning rate $\epsilon$

Prototype Update

new input vector:

$\forall r \in \{1,...,n\} \\ \Delta n_r.p = \epsilon \cdot h(n_r.g,h_s.g) \cdot (x-n_r.p)$

Grid distance function ( )

$h(a,b) = e^{-\frac{(a-b)^2}{s\sigma^2}}$

Constructor & Destructor Documentation

◆ SOM()

icl::math::SOM::SOM	(	unsigned int	dataDim,
		const std::vector< unsigned int > &	dims,
		const std::vector< utils::Range< float > > &	prototypeBounds,
		float	epsilon = `0.1`,
		float	sigma = `1`
	)

create a new SOM with given data dimension and grid dimensions.

Parameters

dataDim	dimension of data elements and prototype vectors
dims	grid dimension array. dims.size() defines the SOM'S grid dimension. dims[i] defines the cell count for dimension i. E.g. to create a 2D 40x40 SOM grid dims has to be $\{40,40\}$ .
prototypeBounds	Internals for new randomly created prototype vector elements. prototypeBounds.size() must be equal to dataDim. E.g. to initialize a SOM with prototypes randomly distributed in a dataDim-D unity cube $[-1,1]^{dataDim}$ , prototypeBounds must be $\{(-1,1),(-1,1),...\}$ (where the ranges are written as tuples).
epsilon	initial learning rate
sigma	initial standard deviation for the grid distance function

◆ ~SOM()

icl::math::SOM::~SOM ( )

Destructor.

Member Function Documentation

◆ getDataDim()

unsigned int icl::math::SOM::getDataDim ( ) const

inline

returns the data dimension

◆ getDimensions()

const std::vector<unsigned int> icl::math::SOM::getDimensions ( ) const

inline

returns the SOM's grid dimension (e.g. 20x40 for a 2D SOM)

◆ getNeuron() [1/2]

const Neuron& icl::math::SOM::getNeuron ( const std::vector< int > & dims ) const

returns a neuron at given grid location (const)

◆ getNeuron() [2/2]

Neuron& icl::math::SOM::getNeuron ( const std::vector< int > & dims )

returns a neuron at given grid location

◆ getNeurons() [1/2]

const std::vector<Neuron>& icl::math::SOM::getNeurons ( ) const

returns the neuron set (const)

◆ getNeurons() [2/2]

std::vector<Neuron>& icl::math::SOM::getNeurons ( )

returns the neuron set

◆ getSomDim()

unsigned int icl::math::SOM::getSomDim ( ) const

inline

returns the SOM dimension (e.g. 2D, 3D ..)

◆ getWinner() [1/2]

const Neuron& icl::math::SOM::getWinner ( const float * input ) const

returns a reference of the winner neuron dependent on a given input vector (const)

◆ getWinner() [2/2]

Neuron& icl::math::SOM::getWinner ( const float * input )

returns a reference of the winner neuron dependent on a given input vector

◆ setEpsilon()

void icl::math::SOM::setEpsilon ( float epsilon )

sets the current learning rate epsilon

◆ setSigma()

void icl::math::SOM::setSigma ( float sigma )

sets the current grid distance function standard deviation

◆ train()

void icl::math::SOM::train ( const float * input )

trains the net using given input vector and current parameters epsilon and sigma

Member Data Documentation

◆ m_fEpsilon

float icl::math::SOM::m_fEpsilon

protected

learning rate

◆ m_fSigma

float icl::math::SOM::m_fSigma

protected

standard deviation for the grid distance function

◆ m_uiDataDim

unsigned int icl::math::SOM::m_uiDataDim

protected

internal data dimension variable

◆ m_uiSomDim

unsigned int icl::math::SOM::m_uiSomDim

protected

internal SOM dimension variable ( = m_vecPrototypeBounds.size() = m_vecDimensions.size() )

◆ m_vecDimensions

std::vector<unsigned int> icl::math::SOM::m_vecDimensions

protected

internal grid dimensions

◆ m_vecDimOffsets

std::vector<unsigned int> icl::math::SOM::m_vecDimOffsets

protected

internal utility offset vector for each dimension

◆ m_vecNeurons

std::vector<Neuron> icl::math::SOM::m_vecNeurons

protected

set of neurons

◆ m_vecPrototypeBounds

std::vector<utils::Range<float> > icl::math::SOM::m_vecPrototypeBounds

protected

internal bounds for prototype ranges (todo: is it necessary to store them ?)

The documentation for this class was generated from the following file:

/Users/alneuman/vm/icl/ICLMath/src/ICLMath/SOM.h

Classes

Public Member Functions

Protected Attributes

Detailed Description

Nomenclature

Prototype Update

Grid distance function ( )

Constructor & Destructor Documentation

◆ SOM()

◆ ~SOM()

Member Function Documentation

◆ getDataDim()

◆ getDimensions()

◆ getNeuron() [1/2]

◆ getNeuron() [2/2]

◆ getNeurons() [1/2]

◆ getNeurons() [2/2]

◆ getSomDim()

◆ getWinner() [1/2]

◆ getWinner() [2/2]

◆ setEpsilon()

◆ setSigma()

◆ train()

Member Data Documentation

◆ m_fEpsilon

◆ m_fSigma

◆ m_uiDataDim

◆ m_uiSomDim

◆ m_vecDimensions

◆ m_vecDimOffsets

◆ m_vecNeurons

◆ m_vecPrototypeBounds