A layer of Gaussian neurons. More...
#include <shark/Unsupervised/RBM/Neuronlayers/GaussianLayer.h>
Public Types | |
typedef RealSpace | StateSpace |
the bias terms associated with the neurons More... | |
typedef RealVector | SufficientStatistics |
The sufficient statistics for the Guassian Layer stores the mean of the neuron and the inverse temperature. More... | |
typedef Batch< SufficientStatistics >::type | StatisticsBatch |
Sufficient statistics of a batch of data. More... | |
Public Types inherited from shark::IParameterizable<> | |
typedef RealVector | ParameterVectorType |
Public Member Functions | |
const RealVector & | bias () const |
Returns the bias values of the units. More... | |
RealVector & | bias () |
Returns the bias values of the units. More... | |
void | resize (std::size_t newSize) |
Resizes this neuron layer. More... | |
std::size_t | size () const |
Returns the number of neurons of this layer. More... | |
template<class Input , class BetaVector > | |
void | sufficientStatistics (Input const &input, StatisticsBatch &statistics, BetaVector const &beta) const |
Takes the input of the neuron and estimates the expectation of the response of the neuron. More... | |
template<class Matrix , class Rng > | |
void | sample (StatisticsBatch const &statistics, Matrix &state, double alpha, Rng &rng) const |
Given a the precomputed statistics (the mean of the Gaussian), the elements of the vector are sampled. This happens either with Gibbs-Sampling or Flip-the-State sampling. For alpha= 0 gibbs sampling is performed. That is the next state for neuron i is directly taken from the conditional distribution of the i-th neuron. In the case of alpha=1, flip-the-state sampling is performed, which takes the last state into account and tries to do deterministically jump into states with higher probability. THIS IS NOT IMPLEMENTED YET and alpha is ignored! More... | |
template<class Matrix > | |
RealVector | logProbability (StatisticsBatch const &statistics, Matrix const &state) const |
Computes the log of the probability of the given states in the conditional distribution. More... | |
template<class Matrix > | |
Matrix const & | phi (Matrix const &state) const |
Transforms the current state of the neurons for the multiplication with the weight matrix of the RBM, i.e. calculates the value of the phi-function used in the interaction term. In the case of Gaussian neurons the phi-function is just the identity. More... | |
RealMatrix const & | expectedPhiValue (StatisticsBatch const &statistics) const |
Returns the expectation of the phi-function. More... | |
RealMatrix const & | mean (StatisticsBatch const &statistics) const |
Returns the mean given the state of the connected layer, i.e. in this case the mean of the Gaussian. More... | |
template<class Matrix , class BetaVector > | |
RealVector | energyTerm (Matrix const &state, BetaVector const &beta) const |
The energy term this neuron adds to the energy function for a batch of inputs. More... | |
template<class Input > | |
double | logMarginalize (const Input &inputs, double beta) const |
Sums over all possible values of the terms of the energy function which depend on the this layer and returns the logarithmic result. More... | |
template<class Vector , class SampleBatch > | |
void | expectedParameterDerivative (Vector &derivative, SampleBatch const &samples) const |
Calculates the expectation of the derivatives of the energy term of this neuron layer with respect to it's parameters - the bias weights. The expectation is taken with respect to the conditional probability distribution of the layer given the state of the connected layer. More... | |
template<class Vector , class SampleBatch , class Vector2 > | |
void | expectedParameterDerivative (Vector &derivative, SampleBatch const &samples, Vector2 const &weights) const |
template<class Vector , class SampleBatch > | |
void | parameterDerivative (Vector &derivative, SampleBatch const &samples) const |
Calculates the derivatives of the energy term of this neuron layer with respect to it's parameters - the bias weights. More... | |
template<class Vector , class SampleBatch , class WeightVector > | |
void | parameterDerivative (Vector &derivative, SampleBatch const &samples, WeightVector const &weights) const |
Calculates the derivatives of the energy term of this neuron layer with respect to it's parameters - the bias weights. More... | |
RealVector | parameterVector () const |
Returns the vector with the parameters associated with the neurons in the layer. More... | |
void | setParameterVector (RealVector const &newParameters) |
Returns the vector with the parameters associated with the neurons in the layer. More... | |
std::size_t | numberOfParameters () const |
Returns the number of the parameters associated with the neurons in the layer. More... | |
void | read (InArchive &archive) |
Reads the bias parameters from an archive. More... | |
void | write (OutArchive &archive) const |
Writes the bias parameters to an archive. More... | |
Public Member Functions inherited from shark::ISerializable | |
virtual | ~ISerializable () |
Virtual d'tor. More... | |
void | load (InArchive &archive, unsigned int version) |
Versioned loading of components, calls read(...). More... | |
void | save (OutArchive &archive, unsigned int version) const |
Versioned storing of components, calls write(...). More... | |
BOOST_SERIALIZATION_SPLIT_MEMBER () | |
Public Member Functions inherited from shark::IParameterizable<> | |
virtual | ~IParameterizable () |
A layer of Gaussian neurons.
For a Gaussian neuron/variable the conditional probability distribution of the state of the variable given the state of the other layer is given by a Gaussian distribution with the input of the neuron as mean and unit variance.
Definition at line 48 of file GaussianLayer.h.
the bias terms associated with the neurons
the state space of this neuron is binary
Definition at line 53 of file GaussianLayer.h.
typedef Batch<SufficientStatistics>::type shark::GaussianLayer::StatisticsBatch |
Sufficient statistics of a batch of data.
Definition at line 58 of file GaussianLayer.h.
typedef RealVector shark::GaussianLayer::SufficientStatistics |
The sufficient statistics for the Guassian Layer stores the mean of the neuron and the inverse temperature.
Definition at line 56 of file GaussianLayer.h.
|
inline |
Returns the bias values of the units.
Definition at line 61 of file GaussianLayer.h.
|
inline |
Returns the bias values of the units.
Definition at line 65 of file GaussianLayer.h.
|
inline |
The energy term this neuron adds to the energy function for a batch of inputs.
state | the state of the neuron layer |
beta | the inverse temperature of the i-th state |
Definition at line 179 of file GaussianLayer.h.
References shark::batchSize(), size(), and SIZE_CHECK.
|
inline |
Calculates the expectation of the derivatives of the energy term of this neuron layer with respect to it's parameters - the bias weights. The expectation is taken with respect to the conditional probability distribution of the layer given the state of the connected layer.
This function takes a batch of samples and extracts the required informations out of it.
derivative | the derivative with respect to the parameters, the result is added on top of it to accumulate derivatives |
samples | the samples from which the informations can be extracted |
Definition at line 229 of file GaussianLayer.h.
References size(), and SIZE_CHECK.
|
inline |
Definition at line 235 of file GaussianLayer.h.
References size(), and SIZE_CHECK.
|
inline |
Returns the expectation of the phi-function.
statistics | the sufficient statistics (the mean of the distribution). |
Definition at line 161 of file GaussianLayer.h.
References size(), and SIZE_CHECK.
|
inline |
Sums over all possible values of the terms of the energy function which depend on the this layer and returns the logarithmic result.
This function is called by Energy when the unnormalized marginal probability of the connected layer is to be computed. This function calculates the part which depends on the neurons which are to be marginalized out. (In the case of the binary hidden neuron, this is the term \( \log \sum_h e^{\vec h^T W \vec v+ \vec h^T \vec c} \)). The rest is calculated by the energy function. In the general case of a hidden layer, this function calculates \( \log \int_h e^(\phi_h(\vec h)^T W \phi_v(\vec v)+f_h(\vec h) ) \) where f_h is the energy term of this.
inputs | the inputs of the neurons they get from the other layer |
beta | the inverse temperature of the RBM |
Definition at line 209 of file GaussianLayer.h.
References beta, size(), SIZE_CHECK, shark::sqr(), and shark::SQRT_2_PI.
|
inline |
Computes the log of the probability of the given states in the conditional distribution.
Currently it is only possible to compute the case with alpha=0
statistics | the statistics of the conditional distribution |
state | the state to check |
Definition at line 132 of file GaussianLayer.h.
References size(), SIZE_CHECK, and shark::sqr().
|
inline |
Returns the mean given the state of the connected layer, i.e. in this case the mean of the Gaussian.
statistics | the sufficient statistics of the layer for a whole batch |
Definition at line 168 of file GaussianLayer.h.
References size(), and SIZE_CHECK.
|
inlinevirtual |
Returns the number of the parameters associated with the neurons in the layer.
Reimplemented from shark::IParameterizable<>.
Definition at line 274 of file GaussianLayer.h.
References size().
|
inline |
Calculates the derivatives of the energy term of this neuron layer with respect to it's parameters - the bias weights.
This function takes a batch of samples and extracts the required informations out of it.
derivative | the derivative with respect to the parameters, the result is added on top of it to accumulate derivatives |
samples | the sample from which the informations can be extracted |
Definition at line 246 of file GaussianLayer.h.
References size(), and SIZE_CHECK.
|
inline |
Calculates the derivatives of the energy term of this neuron layer with respect to it's parameters - the bias weights.
This function takes a batch of samples and calculates a weighted derivative
derivative | the derivative with respect to the parameters, the result is added on top of it to accumulate derivatives |
samples | the sample from which the informations can be extracted |
weights | the weights for the single sample derivatives |
Definition at line 258 of file GaussianLayer.h.
References size(), and SIZE_CHECK.
|
inlinevirtual |
Returns the vector with the parameters associated with the neurons in the layer.
Reimplemented from shark::IParameterizable<>.
Definition at line 264 of file GaussianLayer.h.
|
inline |
Transforms the current state of the neurons for the multiplication with the weight matrix of the RBM, i.e. calculates the value of the phi-function used in the interaction term. In the case of Gaussian neurons the phi-function is just the identity.
state | the state matrix of the neuron layer |
Definition at line 153 of file GaussianLayer.h.
References size(), and SIZE_CHECK.
|
inlinevirtual |
Reads the bias parameters from an archive.
Reimplemented from shark::ISerializable.
Definition at line 279 of file GaussianLayer.h.
|
inline |
Resizes this neuron layer.
newSize | number of neurons in the layer |
Definition at line 72 of file GaussianLayer.h.
|
inline |
Given a the precomputed statistics (the mean of the Gaussian), the elements of the vector are sampled. This happens either with Gibbs-Sampling or Flip-the-State sampling. For alpha= 0 gibbs sampling is performed. That is the next state for neuron i is directly taken from the conditional distribution of the i-th neuron. In the case of alpha=1, flip-the-state sampling is performed, which takes the last state into account and tries to do deterministically jump into states with higher probability. THIS IS NOT IMPLEMENTED YET and alpha is ignored!
statistics | sufficient statistics containing the mean of the conditional Gaussian distribution of the neurons |
state | the state matrix that will hold the sampled states |
alpha | factor changing from gibbs to flip-the state sampling. 0<=alpha<=1 |
rng | the random number generator used for sampling |
Definition at line 110 of file GaussianLayer.h.
References shark::random::gauss(), SHARK_CRITICAL_REGION, size(), and SIZE_CHECK.
|
inlinevirtual |
Returns the vector with the parameters associated with the neurons in the layer.
Reimplemented from shark::IParameterizable<>.
Definition at line 269 of file GaussianLayer.h.
|
inline |
Returns the number of neurons of this layer.
Definition at line 77 of file GaussianLayer.h.
Referenced by energyTerm(), expectedParameterDerivative(), expectedPhiValue(), logMarginalize(), logProbability(), mean(), numberOfParameters(), parameterDerivative(), phi(), sample(), and sufficientStatistics().
|
inline |
Takes the input of the neuron and estimates the expectation of the response of the neuron.
input | the batch of inputs of the neuron |
statistics | sufficient statistics containing the mean of the resulting Gaussian distribution |
beta | the inverse Temperature of the RBM (typically 1) for the whole batch |
Definition at line 87 of file GaussianLayer.h.
References size(), and SIZE_CHECK.
|
inlinevirtual |
Writes the bias parameters to an archive.
Reimplemented from shark::ISerializable.
Definition at line 283 of file GaussianLayer.h.