HopfieldBase Class Reference

A Hopfield network - content addressable memory and energy. More...

#include <HopfieldNetwork.h>

Inheritance diagram for HopfieldBase:

[legend]Collaboration diagram for HopfieldBase:

[legend]List of all members.

Public Member Functions
virtual real	getWeight (uint from, uint to) const=0
virtual void	setWeight (uint from, uint to, real weight)=0
	HopfieldBase (uint size, bool bias=false, bool bipolar=true)
	Creates a Hopfield network, which can store binary patterns.
virtual	~HopfieldBase ()
virtual void	save (const std::string &filename)
	NOT IMPLEMENTED by default.
virtual real	getEnergy ()
	Returns the energy at the current state of the network.
virtual real	getEnergy (int pattern[])
	Returns the energy of the supplied pattern.
virtual uint	getSize () const
	Returns the size of the patterns that can be stored (i.e., length of the pattern vector).
virtual void	step ()
	Steps the network in time.
virtual int	getTime ()
	The number of steps the network has taken since the last call to getOutput.
void	randomize ()
	set random activations
virtual const char *	getClassName () const
	Returns "HopfieldNetwork".
virtual real	getBias (uint i)
	Get the bias at the given neuron.
virtual void	setBias (uint i, real bias)
	Sets the bias of the given neuron.
virtual void	setInput (int pattern[])
	Sets the initial output of the neurons to the values in the given array and resets time to 0.
virtual void	setInput (const std::vector< int > &pattern)
	Sets the initial output of the neuron to the values in the given vector of integers and sets time to 0.
virtual void	setInput (const Vector &pattern)
	Sets the initial output of the neurons to the values in the given vector and sets time to 0.
virtual Vector	getOutput () const
	Returns the current outputs of the neurons.
std::vector< int >	getOutputInt () const
	getOutput clamped to {0,1} or {-1.1}
virtual Vector	getNextOutput ()
	Steps the network through one unit of time and returns the output.
virtual Vector	getOutput (const Vector &input)
	The HopfieldNetwork generally deals with vectors of integers only because the neurons are binary.
virtual bool	propagate (int pattern[], uint timeout)
	Sets the initial input of the network and then steps through time, returning true if the network converges before the given timeout.
virtual bool	propagate (const std::vector< int > &pattern, uint timeout)
	Sets the initial input of the network and then steps through time, returning true if the network converges before the given timeout.
virtual bool	HopfieldBase::propagate (const Vector &pattern, uint timeout)
	the save as discrete version, but may have problems w/ convergence (real values must stand still, not only the ints
Protected Types
typedef uint	Nid
	neuron interface
Protected Member Functions
	HopfieldBase ()
int	real2int (real r) const
	conv to discrete value acc to _bipolar
real	_getBias (Nid i)
void	_setBias (Nid i, real bias)
void	_setNeuronOutput (Nid n, real a)
real	_getNeuronOutput (Nid n) const
Protected Attributes
bool	_bipolar
	true if patterns are bipolar (ie, elements are -1 or 1), false if not (elements are 0 or 1)
int	_nPatterns
	The number of patterns that have been given to the network to be stored.
int	_time
annie::HopfieldBase::NData	_neurons

---

Detailed Description

A Hopfield network - content addressable memory and energy.

This class is an abstract base for real implementations of Hopfield

• either use HopfieldNewtork
• or derive HopfieldBase and define your own getWeight function (this allows to save space and use larger number of neurons - see examples/hdither.cc)

The Hopfield network is considered a landmark in the field of neural computing. This model can store a number of patterns and retrieve the correct pattern starting from a noisy or incomplete pattern so that a network can be trained to identify patterns or images. Hopfield proposed this model as a model of biological memory (often called "content addressable memory"), similar to the way a human brain recognizes an image based on the past experience or learning.

Hopfield also introduced another important idea by establishing a new connection between computation and physics with the concept of energy in his model and showed that computation can be viewed as a transition to lower energy states.

OPT: it would be faster to attach weights by a template instead of inerhitance

---

Member Typedef Documentation

typedef uint Nid [protected]

neuron interface

---

Constructor & Destructor Documentation

HopfieldBase (   )  [inline, protected]

HopfieldBase (  uint  size, bool  bias = false, bool  bipolar = true )

Creates a Hopfield network, which can store binary patterns. Diagonal elments of the weight matrix will be zero. Parameters: size  The size/number of bits of/in the patterns to be stored bias  true if the neurons are to have a bias, false otherwise bipolar  true if patterns are bipolar (elements are -1 or 1), false if not (elements are 0 or 1).

virtual ~HopfieldBase (   )  [inline, virtual]

---

Member Function Documentation

real _getBias (  Nid  i  )  [protected]

real _getNeuronOutput (  Nid  n  )  const [inline, protected]

void _setBias (  Nid  i, real  bias )  [protected]

void _setNeuronOutput (  Nid  n, real  a )  [inline, protected]

virtual real getBias (  uint  i  )  [virtual]

Get the bias at the given neuron. Parameters: i  The neuron whose bias is wanted Returns: The bias

virtual const char* getClassName (   )  const [virtual]

Returns "HopfieldNetwork". Implements Network.

virtual real getEnergy (  int  pattern[]  )  [virtual]

Returns the energy of the supplied pattern. Energy of a pattern depends on the weight matrix

virtual real getEnergy (   )  [virtual]

Returns the energy at the current state of the network.

virtual Vector getNextOutput (   )  [virtual]

Steps the network through one unit of time and returns the output. Returns: A vector v, where v[i] = output of the ith neuron

virtual Vector getOutput (  const Vector &  input  )  [virtual]

The HopfieldNetwork generally deals with vectors of integers only because the neurons are binary. But since this class implements the Network class, it must provide this function as well, which returns the same thing as setInput() following by getOutput() would, except that the neuron's output will be of type real instead of int Implements Network.

virtual Vector getOutput (   )  const [virtual]

Returns the current outputs of the neurons. Returns: A vector v, where v[i] = current output at the ith neuron

std::vector<int> getOutputInt (   )  const

getOutput clamped to {0,1} or {-1.1}

virtual uint getSize (   )  const [virtual]

Returns the size of the patterns that can be stored (i.e., length of the pattern vector).

virtual int getTime (   )  [virtual]

The number of steps the network has taken since the last call to getOutput. See also: getNextOutput getOutput

virtual real getWeight (  uint  from, uint  to )  const [pure virtual]

Implemented in HopfieldNetwork.

virtual bool HopfieldBase::propagate (  const Vector &  pattern, uint  timeout )  [virtual]

the save as discrete version, but may have problems w/ convergence (real values must stand still, not only the ints

virtual bool propagate (  const std::vector< int > &  pattern, uint  timeout )  [virtual]

Sets the initial input of the network and then steps through time, returning true if the network converges before the given timeout. Parameters: pattern  The initial output of the network timeout  The maximum number of steps the network should go through Returns: true if the pattern converges within the timeout, false otherwise

virtual bool propagate (  int  pattern[], uint  timeout )  [virtual]

Sets the initial input of the network and then steps through time, returning true if the network converges before the given timeout. Parameters: pattern  The initial output of the network timeout  The maximum number of steps the network should go through Returns: true if the pattern converges within the timeout, false otherwise

void randomize (   )

set random activations

int real2int (  real  r  )  const [protected]

conv to discrete value acc to _bipolar

virtual void save (  const std::string &  filename  )  [inline, virtual]

NOT IMPLEMENTED by default. Implements Network. Reimplemented in HopfieldNetwork.

virtual void setBias (  uint  i, real  bias )  [virtual]

Sets the bias of the given neuron. Parameters: i  The neuron whose bias is to be set bias  The bias value to give the ith neuron

virtual void setInput (  const Vector &  pattern  )  [virtual]

Sets the initial output of the neurons to the values in the given vector and sets time to 0.

virtual void setInput (  const std::vector< int > &  pattern  )  [virtual]

Sets the initial output of the neuron to the values in the given vector of integers and sets time to 0. Parameters: pattern  A vector of integers containing the initial outputs of the neurons. The vector must contain getSize() elements, each -1 or 1 (if bipolar) or 0 or 1 if not.

virtual void setInput (  int  pattern[]  )  [virtual]

Sets the initial output of the neurons to the values in the given array and resets time to 0. Parameters: pattern  An array containing the initial outputs of the neurons. The array must contain getSize() elements

virtual void setWeight (  uint  from, uint  to, real  weight )  [pure virtual]

Implemented in HopfieldNetwork.

virtual void step (   )  [virtual]

Steps the network in time.

---

Member Data Documentation

bool _bipolar [protected]

true if patterns are bipolar (ie, elements are -1 or 1), false if not (elements are 0 or 1)

struct annie::HopfieldBase::NData _neurons [protected]

int _nPatterns [protected]

The number of patterns that have been given to the network to be stored.

int _time [protected]

---

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

• HopfieldNetwork.h

---