Deep::SimplePCLayer Class Reference

Predictive Coding layer without precision weighting – the simpler, unweighted-energy counterpart to DiscriminativePCLayer. More...

#include <SimplePCLayer.h>

Inheritance diagram for Deep::SimplePCLayer:

Public Member Functions
	SimplePCLayer (size_t size, size_t nextSize, size_t batchSize=1, float learningRate=1e-6f, float inferenceRate=0.1f, float lmbda=1e-2f, void(*act)(float *, size_t)=relu, void(*dAct)(float *, size_t, bool)=dRelu, IComputeBackend *backend=nullptr)
	Constructor for a SimplePCLayer.
	SimplePCLayer (size_t size, size_t nextSize, size_t batchSize=1, float learningRate=1e-6f, float inferenceRate=0.1f, float lmbda=1e-2f, ActivationType aType=ActivationType::RELU, ActivationType dType=ActivationType::dRELU, IComputeBackend *backend=nullptr)
	Constructor for a SimplePCLayer, using a named ActivationType instead of raw function pointers.
float	CalculateState (bool needEnergy=true) noexcept
	Calculates the total network energy state, and this layer's outgoing prediction (mu).
float	CalculateState () noexcept override
	Calculates the internal state of the layer.
void	UpdateState () noexcept override
	Computes the state derivatives for inference.
void	UpdateWeights () noexcept override
	Computes weight updates via gradient descent, with L2 weight decay.
void	Flush () noexcept override
	No-op; exists for Layer interface conformance.
void	ClampState (const std::vector< float > &inputData) noexcept
	Clamps this layer's beliefs to externally-provided data, preventing them from being updated by UpdateState().
void	UnclampState () noexcept
	Releases a previous ClampState() call, allowing this layer's beliefs to update normally again.
bool	IsClamped () const noexcept
	Whether ClampState() is currently active on this layer.
float *	GetBeliefs () noexcept override
	Returns this layer's belief buffer.
const float *	GetMu () const noexcept
	Returns a prediction buffer exposed specifically for forward-projection initialization (seeding the next layer's z from a genuine, current-weights forward pass, rather than zero-init). nullptr for a terminal layer (nextSize=0).
const float *	GetErrors () const noexcept override
	Returns this layer's prediction-error buffer.
size_t	GetInputSize () const noexcept override
	Returns this layer's own belief size.
size_t	GetOutputSize () const noexcept override
	Returns this layer's outgoing prediction size (0 for a terminal layer).
size_t	GetBatchSize () const noexcept override
	Returns the batch size this layer was constructed with.
const float *	GetWeights () const noexcept
	Returns this layer's weight buffer.
float *	GetWeights () noexcept
	Returns this layer's weight buffer.
const float *	GetBiases () const noexcept
	Returns this layer's bias buffer.
float *	GetBiases () noexcept
	Returns this layer's bias buffer.
float	GetLearningRate () const noexcept
	Returns the learning rate used for weight updates.
float	GetInferenceRate () const noexcept
	Returns the inference rate (Euler integration step size).
float	GetLambda () const noexcept
	Returns the weight-decay (L2 regularization) coefficient.
void	SetLearningRate (float lr) noexcept
	Sets the learning rate used for weight updates.
void	SetInferenceRate (float ir) noexcept
	Sets the inference rate (Euler integration step size).
void	SetLambda (float l) noexcept
	Sets the weight-decay (L2 regularization) coefficient.
void	SetOptimizer (const OptimizerType o) noexcept
	Selects the optimizer used for weight updates (SGD or Adam). Adam's moment buffers are lazily allocated on first use.
void	SetMuCacheThreshold (float threshold) noexcept
	Sets the mu-cache staleness threshold: mu is recomputed only when ||z - z_at_last_recompute|| / (||z_at_last_recompute||.
float	GetMuCacheThreshold () const noexcept
	Returns the current mu-cache staleness threshold – see SetMuCacheThreshold() above.
void	ComputeMuOnly () noexcept
	Computes only mu (forward prediction), skipping error/energy entirely. Extracted from CalculateState() for callers (like ProjectForward()) that don't need the discarded error/energy values. Computes zF=phi(z) internally and uses it for the GEMM, matching the activate-before-transform convention (mu=W@phi(z)+b).
void	SetLayerAbove (SimplePCLayer *above) noexcept
	Sets the layer immediately above this one in the network.
void	SetLayerBelow (SimplePCLayer *below) noexcept
	Sets the layer immediately below this one in the network.
void	ResetState () noexcept
	Resets this layer's beliefs/errors back to their initial values, without touching learned weights.
const SimplePCLayer &	GetLayerAbove () const noexcept
	Returns the layer immediately above this one.
const SimplePCLayer &	GetLayerBelow () const noexcept
	Returns the layer immediately below this one.
void	RandomizeWeights (std::mt19937 &twister, const char *distribution="normal") noexcept
	Randomizes this layer's weights (and biases) in place.
ActivationType	GetActivationType () const noexcept
	Returns this layer's configured activation type.
ActivationType	GetDerivativeType () const noexcept
	Returns this layer's configured activation-derivative type.
size_t	GetRequiredFloats () const noexcept
	Computes the total number of floats this layer requires from a MemoryArena (weights, biases, beliefs, errors, scratch buffers, and, if Adam is selected, its moment buffers).
template<typename ArenaT>
void	BindMemory (ArenaT &arena)
	Binds this layer's weight/state/scratch buffers into the supplied arena. Must be called before any other operation. Templated so either MemoryArena (CPU) or DeviceMemoryArena (GPU) can be bound, resolved entirely at compile time – see the .cpp's explicit instantiations for the two concrete types actually used.
Public Member Functions inherited from Deep::Layer
virtual	~Layer ()=default
virtual std::map< std::string, TensorDescriptor >	GetStateDict () const
	Returns a dictionary of this.

Friends
class	SimplePCNDiagnostics

Additional Inherited Members
Protected Attributes inherited from Deep::Layer
size_t	size
	Size of input.
size_t	nextSize
	Size of output.

Detailed Description

Predictive Coding layer without precision weighting – the simpler, unweighted-energy counterpart to DiscriminativePCLayer.

Constructor & Destructor Documentation

SimplePCLayer() [1/2]

Deep::SimplePCLayer::SimplePCLayer	(	size_t	size,
		size_t	nextSize,
		size_t	batchSize = 1,
		float	learningRate = 1e-6f,
		float	inferenceRate = 0.1f,
		float	lmbda = 1e-2f,
		void(*	act )(float *, size_t) = relu,
		void(*	dAct )(float *, size_t, bool) = dRelu,
		IComputeBackend *	backend = nullptr )

Constructor for a SimplePCLayer.

Parameters

size	Size of this layer's own belief (z)
nextSize	Size of the layer above's belief (this layer's outgoing prediction target); 0 marks a terminal layer
batchSize	Batch size
learningRate	Learning rate for weight updates
inferenceRate	Inference rate (Euler integration step size)
lmbda	Weight decay (L2 regularization) coefficient
act	Activation function
dAct	Derivative of activation function
backend	Compute backend to route all math through. Defaults to nullptr, in which case this layer constructs and owns its own CPUBackend internally – existing callers don't need to change anything. Pass a real backend (owned elsewhere, e.g. by the network) to run this layer on GPU.

SimplePCLayer() [2/2]

Deep::SimplePCLayer::SimplePCLayer	(	size_t	size,
		size_t	nextSize,
		size_t	batchSize = 1,
		float	learningRate = 1e-6f,
		float	inferenceRate = 0.1f,
		float	lmbda = 1e-2f,
		ActivationType	aType = ActivationType::RELU,
		ActivationType	dType = ActivationType::dRELU,
		IComputeBackend *	backend = nullptr )

Constructor for a SimplePCLayer, using a named ActivationType instead of raw function pointers.

Parameters

size	Size of this layer's own belief (z)
nextSize	Size of the layer above's belief (this layer's outgoing prediction target); 0 marks a terminal layer
batchSize	Batch size
learningRate	Learning rate for weight updates
inferenceRate	Inference rate (Euler integration step size)
lmbda	Weight decay (L2 regularization) coefficient
aType	Activation type
dType	Activation derivative type
backend	Compute backend to route all math through. See the other constructor's doc for the default-nullptr behavior.

Member Function Documentation

BindMemory()

template<typename ArenaT>

template void Deep::SimplePCLayer::BindMemory< DeviceMemoryArena >

(

ArenaT &

arena

)

Binds this layer's weight/state/scratch buffers into the supplied arena. Must be called before any other operation. Templated so either MemoryArena (CPU) or DeviceMemoryArena (GPU) can be bound, resolved entirely at compile time – see the .cpp's explicit instantiations for the two concrete types actually used.

Parameters

arena

The arena to bind into.

CalculateState() [1/2]

float Deep::SimplePCLayer::CalculateState ( )

inlineoverridevirtualnoexcept

Calculates the internal state of the layer.

Returns

energy used

Implements Deep::Layer.

CalculateState() [2/2]

float Deep::SimplePCLayer::CalculateState ( bool needEnergy = true )

noexcept

Calculates the total network energy state, and this layer's outgoing prediction (mu).

\[E = \sum_l 1/2 ||z^{(l)} - \mu^{(l)}||^2 \]

mu is now computed as mu = W @ phi(z) + b – activation applied to z BEFORE the linear transform, matching ngc-learn's documented convention exactly (was previously mu = phi(W@z+b), activation AFTER the transform). (No precision weighting – see file-level note.)

Parameters

needEnergy

Asks for energy

Returns

This layer's energy contribution at the current state, if asked for.

ClampState()

void Deep::SimplePCLayer::ClampState ( const std::vector< float > & inputData )

noexcept

Clamps this layer's beliefs to externally-provided data, preventing them from being updated by UpdateState().

Parameters

inputData

Flattened input data of length size.

ComputeMuOnly()

void Deep::SimplePCLayer::ComputeMuOnly ( )

noexcept

Computes only mu (forward prediction), skipping error/energy entirely. Extracted from CalculateState() for callers (like ProjectForward()) that don't need the discarded error/energy values. Computes zF=phi(z) internally and uses it for the GEMM, matching the activate-before-transform convention (mu=W@phi(z)+b).

Flush()

void Deep::SimplePCLayer::Flush ( )

inlineoverridevirtualnoexcept

No-op; exists for Layer interface conformance.

Reimplemented from Deep::Layer.

GetActivationType()

ActivationType Deep::SimplePCLayer::GetActivationType ( ) const

inlinenoexcept

Returns this layer's configured activation type.

GetBatchSize()

size_t Deep::SimplePCLayer::GetBatchSize ( ) const

inlineoverridevirtualnoexcept

Returns the batch size this layer was constructed with.

Implements Deep::Layer.

GetBeliefs()

float * Deep::SimplePCLayer::GetBeliefs ( )

inlineoverridevirtualnoexcept

Returns this layer's belief buffer.

Returns

Pointer to this layer's size-length beliefs.

Implements Deep::Layer.

GetBiases() [1/2]

const float * Deep::SimplePCLayer::GetBiases ( ) const

inlinenoexcept

Returns this layer's bias buffer.

GetBiases() [2/2]

float * Deep::SimplePCLayer::GetBiases ( )

inlinenoexcept

Returns this layer's bias buffer.

GetDerivativeType()

ActivationType Deep::SimplePCLayer::GetDerivativeType ( ) const

inlinenoexcept

Returns this layer's configured activation-derivative type.

GetErrors()

const float * Deep::SimplePCLayer::GetErrors ( ) const

inlineoverridevirtualnoexcept

Returns this layer's prediction-error buffer.

Returns

Pointer to this layer's size-length errors.

Implements Deep::Layer.

GetInferenceRate()

float Deep::SimplePCLayer::GetInferenceRate ( ) const

inlinenoexcept

Returns the inference rate (Euler integration step size).

GetInputSize()

size_t Deep::SimplePCLayer::GetInputSize ( ) const

inlineoverridevirtualnoexcept

Returns this layer's own belief size.

Implements Deep::Layer.

GetLambda()

float Deep::SimplePCLayer::GetLambda ( ) const

inlinenoexcept

Returns the weight-decay (L2 regularization) coefficient.

GetLayerAbove()

const SimplePCLayer & Deep::SimplePCLayer::GetLayerAbove ( ) const

inlinenoexcept

Returns the layer immediately above this one.

Warning

Dereferences layerAbove without a null check; only valid if SetLayerAbove() was previously called with a non-null pointer.

GetLayerBelow()

const SimplePCLayer & Deep::SimplePCLayer::GetLayerBelow ( ) const

inlinenoexcept

Returns the layer immediately below this one.

Warning

Dereferences layerBelow without a null check; only valid if SetLayerBelow() was previously called with a non-null pointer.

GetLearningRate()

float Deep::SimplePCLayer::GetLearningRate ( ) const

inlinenoexcept

Returns the learning rate used for weight updates.

GetMu()

const float * Deep::SimplePCLayer::GetMu ( ) const

inlinenoexcept

Returns a prediction buffer exposed specifically for forward-projection initialization (seeding the next layer's z from a genuine, current-weights forward pass, rather than zero-init). nullptr for a terminal layer (nextSize=0).

Returns

Outgoing prediction buffer

GetMuCacheThreshold()

float Deep::SimplePCLayer::GetMuCacheThreshold ( ) const

inlinenoexcept

Returns the current mu-cache staleness threshold – see SetMuCacheThreshold() above.

GetOutputSize()

size_t Deep::SimplePCLayer::GetOutputSize ( ) const

inlineoverridevirtualnoexcept

Returns this layer's outgoing prediction size (0 for a terminal layer).

Implements Deep::Layer.

GetRequiredFloats()

size_t Deep::SimplePCLayer::GetRequiredFloats ( ) const

noexcept

Computes the total number of floats this layer requires from a MemoryArena (weights, biases, beliefs, errors, scratch buffers, and, if Adam is selected, its moment buffers).

Returns

The required float count.

GetWeights() [1/2]

const float * Deep::SimplePCLayer::GetWeights ( ) const

inlinenoexcept

Returns this layer's weight buffer.

GetWeights() [2/2]

float * Deep::SimplePCLayer::GetWeights ( )

inlinenoexcept

Returns this layer's weight buffer.

IsClamped()

bool Deep::SimplePCLayer::IsClamped ( ) const

inlinenoexcept

Whether ClampState() is currently active on this layer.

RandomizeWeights()

void Deep::SimplePCLayer::RandomizeWeights ( std::mt19937 & twister, const char * distribution = "normal" )

noexcept

Randomizes this layer's weights (and biases) in place.

Parameters

twister	The classic Mersenne Twister
distribution	The distribution of the randomization: (normal, uniform)

ResetState()

void Deep::SimplePCLayer::ResetState ( )

noexcept

Resets this layer's beliefs/errors back to their initial values, without touching learned weights.

SetInferenceRate()

void Deep::SimplePCLayer::SetInferenceRate ( float ir )

inlinenoexcept

Sets the inference rate (Euler integration step size).

Parameters

ir	The new inference rate.

SetLambda()

void Deep::SimplePCLayer::SetLambda ( float l )

inlinenoexcept

Sets the weight-decay (L2 regularization) coefficient.

Parameters

l	The new lambda value.

SetLayerAbove()

void Deep::SimplePCLayer::SetLayerAbove ( SimplePCLayer * above )

inlinenoexcept

Sets the layer immediately above this one in the network.

Parameters

above

Pointer to the layer above; may be nullptr for a terminal layer.

SetLayerBelow()

void Deep::SimplePCLayer::SetLayerBelow ( SimplePCLayer * below )

inlinenoexcept

Sets the layer immediately below this one in the network.

Parameters

below

Pointer to the layer below; may be nullptr for the input layer.

SetLearningRate()

void Deep::SimplePCLayer::SetLearningRate ( float lr )

noexcept

Sets the learning rate used for weight updates.

Parameters

lr	The new learning rate.

SetMuCacheThreshold()

void Deep::SimplePCLayer::SetMuCacheThreshold ( float threshold )

inlinenoexcept

Sets the mu-cache staleness threshold: mu is recomputed only when ||z - z_at_last_recompute|| / (||z_at_last_recompute||.

• eps) exceeds this value. threshold=-1 disables caching entirely (default). threshold=0 reproduces the original, EXACT clamped-only behavior (a clamped layer's z never changes, so its ratio is always exactly 0, always below any threshold>=0). threshold>0 extends caching to UNCLAMPED layers too, as a genuine approximation – correctness there means "close enough for real training," not "bit-identical," and needs its own accuracy-impact validation, not just a trajectory diff.

SetOptimizer()

void Deep::SimplePCLayer::SetOptimizer ( const OptimizerType o )

inlinenoexcept

Selects the optimizer used for weight updates (SGD or Adam). Adam's moment buffers are lazily allocated on first use.

Parameters

o	The optimizer type to use.

UnclampState()

void Deep::SimplePCLayer::UnclampState ( )

noexcept

Releases a previous ClampState() call, allowing this layer's beliefs to update normally again.

UpdateState()

void Deep::SimplePCLayer::UpdateState ( )

overridevirtualnoexcept

Computes the state derivatives for inference.

\[\frac{dz^{(l)}}{dt} = -e^{(l)} + \sigma'(z^{(l)}) \odot (W^{(l-1)})^T e^{(l-1)} \]

The derivative multiply now applies AFTER the W transform, using f'(z) (this layer's OWN state derivative), not f'(mu) applied before – matching the activate-before-transform convention.

Implements Deep::Layer.

UpdateWeights()

void Deep::SimplePCLayer::UpdateWeights ( )

overridevirtualnoexcept

Computes weight updates via gradient descent, with L2 weight decay.

Implements Deep::Layer.

Friends And Related Symbol Documentation

SimplePCNDiagnostics

friend class SimplePCNDiagnostics

friend

---

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

• include/deepity/layers/SimplePCLayer.h
• src/layers/SimplePCLayer.cpp