Deep::FullPCNetwork Class Reference

Predictive Coding network combining muPC scaling, optional residual connections, and DKP direct feedback – see the file-level note above for how its defaults reproduce DirectKPPCNetwork exactly. More...

#include <FullPCNetwork.h>

Public Member Functions
	FullPCNetwork (int batchSize, DeviceType device=DeviceType::DEVICE_CPU) noexcept
	Constructs an empty network; add layers via AddLayer(), then Compile() before use.
	~FullPCNetwork ()=default
	FullPCNetwork (const FullPCNetwork &)=delete
FullPCNetwork &	operator= (const FullPCNetwork &)=delete
void	AddLayer (size_t size, size_t nextSize, size_t terminalSize, float lr, float ir, float fl, float lmbda, ActivationType aType, ActivationType dType)
	Adds a layer to the network.
void	SetUseMuPCScaling (bool enabled) noexcept
	Enables muPC-style per-layer forward scaling (Table 1 of the muPC paper). OFF by default. Must be called before Compile() – Compile() is what actually computes and applies each layer's a, once the full architecture (every AddLayer call) is known.
void	SetUseResidualConnections (bool enabled) noexcept
	Enables residual/skip connections on middle hidden layers (muPC's "1-skip" ResNet formulation – see that paper's A.2.4). OFF by default. Requires every middle hidden layer to have the SAME width as its neighbor; Compile() throws std::invalid_argument if that's violated, rather than silently skipping the mismatched connection (a silent skip risks masking a real configuration mistake).
void	RandomizeWeights (std::mt19937 &rng)
	Randomizes the weights (W and Psi) of each layer.
void	ResetState () noexcept
	Resets each layer's state without touching learned weights.
void	Clamp (const std::vector< float > &input)
	Clamps the input to the first layer.
void	ProjectForward () noexcept
	Seeds hidden layers from a genuine forward pass through current weights, instead of zero-init. Call after Clamp(), before the settling loop.
float	CalculateTerminalError () noexcept
	Computes the terminal layer's error/energy against its clamped target. Call AFTER clamping the target onto the terminal layer, and BEFORE DirectFeedbackUpdate().
void	DirectFeedbackUpdate () noexcept
	Runs the DFA phase on every non-terminal layer: perturbs each layer's W using the layer above's Psi and the terminal layer's error.
float	Step (bool computeEnergy=true) noexcept
	Runs one settling step on every layer.
void	UpdateWeights () noexcept
	Applies weight updates to every layer.
void	SetOptimizer (OptimizerType o) noexcept
	Sets the weight optimizer (W) on every layer.
void	SetPsiOptimizer (OptimizerType o) noexcept
	Sets the feedback-weight optimizer (Psi) on every layer.
void	SetLearningRate (float lr) noexcept
	Sets the learning rate (W) on every layer.
void	SetFeedbackRate (float fl) noexcept
	Sets the feedback rate (Psi) on every layer.
FullPCLayer *	GetTerminalLayer () noexcept
	Returns the last layer (nextSize==0, the terminal one).
std::vector< std::unique_ptr< FullPCLayer > > &	GetLayers () noexcept
	Returns every layer in the network, in the order they were added.
const std::vector< std::unique_ptr< FullPCLayer > > &	GetLayers () const noexcept
	const overload of GetLayers() above.
int	GetBatchSize () const noexcept
	Returns the batch size given at construction.
DeviceType	GetDevice () const noexcept
	Returns which device this network's layers run on.
float	TrainStep (const std::vector< float > &x, const std::vector< float > &y, int inferenceSteps=1)
	Full train step: reset, clamp input+target, run all four DKP-PC phases in order, unclamp.
std::vector< float >	Predict (const std::vector< float > &x, int inferenceSteps)
	Forward prediction: clamp input, settle with no DFA perturbation and no target clamped, read the terminal's beliefs.
void	Compile ()
	Loads all layers into one contiguous block of memory, wires layerAbove/layerBelow/terminalLayer across every layer – same as DirectKPPCNetwork::Compile() – PLUS: if useMuPCScaling, computes and applies each layer's a (Table 1) from the full, now-known architecture; if useResidualConnections, applies SetResidual(true) to every middle hidden layer after checking width match (throws std::invalid_argument on mismatch).
void	SetUseIPC (bool enabled) noexcept
	Enables iPC: UpdateWeights() runs every settling step instead of once after settling completes. OFF by default – TrainStep() matches DirectKPPCNetwork's standard, two-phase behavior exactly when this is false.
void	SetUseMomentum (bool enabled, float beta=0.9f) noexcept
	Enables momentum (inertial) settling on every layer. Loops over layers and calls each one's own SetMomentum() – same pattern as SetLearningRate/SetFeedbackRate. OFF by default.
void	SetUseCrossEntropy (bool enabled) noexcept
	Enables softmax cross-entropy energy on the TERMINAL layer only (unlike the other Set* toggles, this does NOT loop over every layer – cross-entropy only ever makes sense on the final output, against a one-hot/class-probability target). OFF by default (plain Gaussian energy everywhere, matching DirectKPPCNetwork).

Detailed Description

Predictive Coding network combining muPC scaling, optional residual connections, and DKP direct feedback – see the file-level note above for how its defaults reproduce DirectKPPCNetwork exactly.

Constructor & Destructor Documentation

FullPCNetwork() [1/2]

Deep::FullPCNetwork::FullPCNetwork ( int batchSize, DeviceType device = DeviceType::DEVICE_CPU )

explicitnoexcept

Constructs an empty network; add layers via AddLayer(), then Compile() before use.

Parameters

batchSize	Fixed batch size for every layer.
device	Which device this network's layers run on.

~FullPCNetwork()

Deep::FullPCNetwork::~FullPCNetwork ( )

default

FullPCNetwork() [2/2]

Deep::FullPCNetwork::FullPCNetwork ( const FullPCNetwork & )

delete

Member Function Documentation

AddLayer()

void Deep::FullPCNetwork::AddLayer	(	size_t	size,
		size_t	nextSize,
		size_t	terminalSize,
		float	lr,
		float	ir,
		float	fl,
		float	lmbda,
		ActivationType	aType,
		ActivationType	dType )

Adds a layer to the network.

Parameters

size	input size
nextSize	output size (0 marks a terminal layer)
terminalSize	the size of the network's final output layer (e.g. 10 for MNIST) – required on every AddLayer call, not inferred, since the true terminal layer isn't known until the whole network has been assembled. Compile() sanity-checks this against the actual last layer's size.
lr	learning rate for W
ir	inference rate (Euler integration step size)
fl	feedback rate (see FullPCLayer's own docs)
lmbda	weight decay (L2 regularization) coefficient, shared between W and Psi
aType	activation type
dType	activation derivative type

CalculateTerminalError()

float Deep::FullPCNetwork::CalculateTerminalError ( )

noexcept

Computes the terminal layer's error/energy against its clamped target. Call AFTER clamping the target onto the terminal layer, and BEFORE DirectFeedbackUpdate().

Returns

The terminal layer's energy contribution.

Clamp()

void Deep::FullPCNetwork::Clamp

(

const std::vector< float > &

input

)

Clamps the input to the first layer.

Parameters

input

Reference to the input vector.

Compile()

void Deep::FullPCNetwork::Compile

(

)

Loads all layers into one contiguous block of memory, wires layerAbove/layerBelow/terminalLayer across every layer – same as DirectKPPCNetwork::Compile() – PLUS: if useMuPCScaling, computes and applies each layer's a (Table 1) from the full, now-known architecture; if useResidualConnections, applies SetResidual(true) to every middle hidden layer after checking width match (throws std::invalid_argument on mismatch).

DirectFeedbackUpdate()

void Deep::FullPCNetwork::DirectFeedbackUpdate ( )

noexcept

Runs the DFA phase on every non-terminal layer: perturbs each layer's W using the layer above's Psi and the terminal layer's error.

GetBatchSize()

int Deep::FullPCNetwork::GetBatchSize ( ) const

inlinenoexcept

Returns the batch size given at construction.

GetDevice()

DeviceType Deep::FullPCNetwork::GetDevice ( ) const

inlinenoexcept

Returns which device this network's layers run on.

GetLayers() [1/2]

const std::vector< std::unique_ptr< FullPCLayer > > & Deep::FullPCNetwork::GetLayers ( ) const

inlinenoexcept

const overload of GetLayers() above.

GetLayers() [2/2]

std::vector< std::unique_ptr< FullPCLayer > > & Deep::FullPCNetwork::GetLayers ( )

inlinenoexcept

Returns every layer in the network, in the order they were added.

GetTerminalLayer()

FullPCLayer * Deep::FullPCNetwork::GetTerminalLayer ( )

inlinenoexcept

Returns the last layer (nextSize==0, the terminal one).

operator=()

FullPCNetwork & Deep::FullPCNetwork::operator= ( const FullPCNetwork & )

delete

Predict()

std::vector< float > Deep::FullPCNetwork::Predict	(	const std::vector< float > &	x,
		int	inferenceSteps )

Forward prediction: clamp input, settle with no DFA perturbation and no target clamped, read the terminal's beliefs.

ProjectForward()

void Deep::FullPCNetwork::ProjectForward ( )

noexcept

Seeds hidden layers from a genuine forward pass through current weights, instead of zero-init. Call after Clamp(), before the settling loop.

RandomizeWeights()

void Deep::FullPCNetwork::RandomizeWeights

(

std::mt19937 &

rng

)

Randomizes the weights (W and Psi) of each layer.

Parameters

rng	The classic Mersenne Twister

ResetState()

void Deep::FullPCNetwork::ResetState ( )

noexcept

Resets each layer's state without touching learned weights.

SetFeedbackRate()

void Deep::FullPCNetwork::SetFeedbackRate ( float fl )

inlinenoexcept

Sets the feedback rate (Psi) on every layer.

SetLearningRate()

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

inlinenoexcept

Sets the learning rate (W) on every layer.

SetOptimizer()

void Deep::FullPCNetwork::SetOptimizer ( OptimizerType o )

inlinenoexcept

Sets the weight optimizer (W) on every layer.

SetPsiOptimizer()

void Deep::FullPCNetwork::SetPsiOptimizer ( OptimizerType o )

inlinenoexcept

Sets the feedback-weight optimizer (Psi) on every layer.

SetUseCrossEntropy()

void Deep::FullPCNetwork::SetUseCrossEntropy ( bool enabled )

inlinenoexcept

Enables softmax cross-entropy energy on the TERMINAL layer only (unlike the other Set* toggles, this does NOT loop over every layer – cross-entropy only ever makes sense on the final output, against a one-hot/class-probability target). OFF by default (plain Gaussian energy everywhere, matching DirectKPPCNetwork).

SetUseIPC()

void Deep::FullPCNetwork::SetUseIPC ( bool enabled )

inlinenoexcept

Enables iPC: UpdateWeights() runs every settling step instead of once after settling completes. OFF by default – TrainStep() matches DirectKPPCNetwork's standard, two-phase behavior exactly when this is false.

See also

Salvatori et al., "Incremental Predictive Coding", https://arxiv.org/abs/2212.00720

SetUseMomentum()

void Deep::FullPCNetwork::SetUseMomentum ( bool enabled, float beta = 0.9f )

inlinenoexcept

Enables momentum (inertial) settling on every layer. Loops over layers and calls each one's own SetMomentum() – same pattern as SetLearningRate/SetFeedbackRate. OFF by default.

SetUseMuPCScaling()

void Deep::FullPCNetwork::SetUseMuPCScaling ( bool enabled )

inlinenoexcept

Enables muPC-style per-layer forward scaling (Table 1 of the muPC paper). OFF by default. Must be called before Compile() – Compile() is what actually computes and applies each layer's a, once the full architecture (every AddLayer call) is known.

SetUseResidualConnections()

void Deep::FullPCNetwork::SetUseResidualConnections ( bool enabled )

inlinenoexcept

Enables residual/skip connections on middle hidden layers (muPC's "1-skip" ResNet formulation – see that paper's A.2.4). OFF by default. Requires every middle hidden layer to have the SAME width as its neighbor; Compile() throws std::invalid_argument if that's violated, rather than silently skipping the mismatched connection (a silent skip risks masking a real configuration mistake).

Step()

float Deep::FullPCNetwork::Step ( bool computeEnergy = true )

noexcept

Runs one settling step on every layer.

Parameters

computeEnergy

asks for energy to be returned

Returns

Total energy, summed from every layer's CalculateState().

TrainStep()

float Deep::FullPCNetwork::TrainStep	(	const std::vector< float > &	x,
		const std::vector< float > &	y,
		int	inferenceSteps = 1 )

Full train step: reset, clamp input+target, run all four DKP-PC phases in order, unclamp.

Parameters

x	Flattened input batch, clamped to the input layer.
y	Flattened target batch, clamped to the terminal layer.
inferenceSteps	Number of settling steps for phase 2.

Returns

The final settling step's total energy, before weight updates.

UpdateWeights()

void Deep::FullPCNetwork::UpdateWeights ( )

noexcept

Applies weight updates to every layer.

---

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

• include/deepity/networks/FullPCNetwork.h
• src/networks/FullPCNetwork.cpp