nest::EpropArchivingNodeRecurrent< hist_shift_required > Class Template Reference

Class implementing an intermediate archiving node model for recurrent node models supporting e-prop plasticity. More...

#include <eprop_archiving_node_recurrent.h>

Inheritance diagram for nest::EpropArchivingNodeRecurrent< hist_shift_required >:

Collaboration diagram for nest::EpropArchivingNodeRecurrent< hist_shift_required >:

Public Types
typedef double(EpropArchivingNodeRecurrent::*	surrogate_gradient_function) (double, double, double, double, double)
	Defines the pointer-to-member function type for the surrogate gradient function.

Public Member Functions
	EpropArchivingNodeRecurrent ()
	Constructs a new EpropArchivingNodeRecurrent object.
	EpropArchivingNodeRecurrent (const EpropArchivingNodeRecurrent &other)
	Constructs an EpropArchivingNodeRecurrent object by copying another EpropArchivingNodeRecurrent object.
surrogate_gradient_function	find_surrogate_gradient (const std::string &surrogate_gradient_function_name)
	Validates and finds surrogate gradient function based on the specified name.
double	compute_piecewise_linear_surrogate_gradient (const double r, const double v_m, const double v_th, const double height, const double width)
	Computes a piecewise-linear function as the surrogate gradient.
double	compute_exponential_surrogate_gradient (const double r, const double v_m, const double v_th, const double height, const double width)
	Computes an exponential function as the surrogate gradient.
double	compute_fast_sigmoid_derivative_surrogate_gradient (const double r, const double v_m, const double v_th, const double height, const double width)
	Computes the derivative of a fast-sigmoid surrogate function as the surrogate gradient.
double	compute_arctan_derivative_surrogate_gradient (const double r, const double v_m, const double v_th, const double height, const double width)
	Computes the derivative of an arctan surrogate function as the surrogate gradient.
void	append_new_eprop_history_entry (const long time_step)
	Creates an entry for the specified time step at the end of the eprop history.
void	write_surrogate_gradient_to_history (const long time_step, const double surrogate_gradient)
	Writes the surrogate gradient to the eprop history entry at the specified time step.
void	write_learning_signal_to_history (const long time_step, const double learning_signal)
	Writes the learning signal to the eprop history entry at the specifed time step.
void	write_firing_rate_reg_to_history (const long t_current_update, const double f_target, const double c_reg)
	Calculates the firing rate regularization for the current update and writes it to a new entry in the firing rate regularization history.
void	write_firing_rate_reg_to_history (const long time_step, const double z, const double f_target, const double kappa_reg, const double c_reg)
	Calculates the current firing rate regularization and writes it to the eprop history at the specified time step.
double	get_firing_rate_reg_history (const long time_step)
	Retrieves the firing rate regularization at the specified time step from the firing rate regularization history.
double	get_learning_signal_from_history (const long time_step)
	Retrieves the learning signal from the eprop history at the specified time step.
void	erase_used_firing_rate_reg_history ()
	Erases the history of the used firing rate regularization history.
void	count_spike ()
	Counts an emitted spike for the firing rate regularization.
void	reset_spike_count ()
	Resets the spike count for the firing rate regularization.
void	get_status (Dictionary &d) const override
	Export properties of the node by setting entries in the status dictionary.
void	set_status (const Dictionary &d) override
	Change properties of the node according to the entries in the dictionary.
Public Member Functions inherited from nest::EpropArchivingNode< HistEntryEpropRecurrent >
	EpropArchivingNode ()
	Constructs a new EpropArchivingNode object.
	EpropArchivingNode (const EpropArchivingNode &n)
	Constructs a new EpropArchivingNode object by copying another EpropArchivingNode object.
void	register_eprop_connection () override
	Registers an eprop connection.
void	initialize_update_history () override
	Initializes the update history.
void	write_update_to_history (const long t_previous_update, const long t_current_update, const bool is_flush_event, const bool previous_was_flush_event) override
	Registers the current update in the update history and deregisters the previous update.
void	require_eprop_history_entry (const typename std::vector< HistEntryEpropRecurrent >::iterator eprop_hist_it, const long time_step) const
	Requires that an e-prop history iterator points to the expected time step.
std::vector< HistEntryEpropUpdate >::iterator	get_update_history (const long time_step)
	Retrieves the update history entry for a specific time step.
std::vector< HistEntryEpropRecurrent >::iterator	get_eprop_history (const long time_step)
	Retrieves the eprop history entry for a specified time step.
void	erase_used_eprop_history () override
	Erases the used eprop history for bsshslm_2020 models.
void	erase_used_eprop_history (const long t_spike, const long t_spike_previous) override
	Erases the used eprop history.
double	get_eprop_history_duration () const
	Retrieves eprop history size.
long	get_eprop_isi_trace_cutoff () const
	Retrieves the eprop ISI trace cutoff.
Public Member Functions inherited from nest::Node
	Node ()
	Node (Node const &)
virtual	~Node ()
virtual Node *	clone () const
	This function creates a new object by calling the derived class' copy constructor and returning its pointer.
virtual bool	has_proxies () const
	Returns true if the node has proxies on remote threads.
virtual bool	supports_urbanczik_archiving () const
	Returns true if the node supports the Urbanczik-Senn plasticity rule.
virtual bool	local_receiver () const
	Returns true if the node only receives events from nodes/devices on the same thread.
virtual bool	one_node_per_process () const
	Returns true if the node exists only once per process, but does not have proxies on remote threads.
virtual bool	is_off_grid () const
	Returns true if the node sends/receives off-grid events.
virtual bool	is_proxy () const
	Returns true if the node is a proxy node.
std::string	get_name () const
	Return class name.
virtual std::string	get_element_type () const
	Return the element type of the node.
size_t	get_node_id () const
	Return global Network ID.
int	get_model_id () const
	Return model ID of the node.
bool	is_frozen () const
	Returns true if node is frozen, i.e., shall not be updated.
bool	node_uses_wfr () const
	Returns true if the node uses the waveform relaxation method.
void	set_node_uses_wfr (const bool)
	Sets node_uses_wfr_ member variable (to be able to set it to "true" for any class derived from Node)
void	init ()
	Initialize node prior to first simulation after node has been created.
virtual void	pre_run_hook ()=0
	Re-calculate dependent parameters of the node.
virtual void	calibrate_time (const TimeConverter &)
	Re-calculate time-based properties of the node.
virtual void	post_run_cleanup ()
	Cleanup node after Run.
virtual void	finalize ()
	Finalize node.
virtual void	update (Time const &, const long, const long)=0
	Advance the state of the node in time through the given interval.
virtual bool	wfr_update (Time const &, const long, const long)
	Advance the state of the node in time through the given interval (see Node::update() for more details).
virtual size_t	send_test_event (Node &receiving_node, size_t receptor_type, synindex syn_id, bool dummy_target)
	Send an event to the receiving_node passed as an argument.
virtual size_t	handles_test_event (SpikeEvent &, size_t receptor_type)
	Check if the node can handle a particular event and receptor type.
virtual size_t	handles_test_event (WeightRecorderEvent &, size_t receptor_type)
virtual size_t	handles_test_event (RateEvent &, size_t receptor_type)
virtual size_t	handles_test_event (DataLoggingRequest &, size_t receptor_type)
virtual size_t	handles_test_event (CurrentEvent &, size_t receptor_type)
virtual size_t	handles_test_event (ConductanceEvent &, size_t receptor_type)
virtual size_t	handles_test_event (DoubleDataEvent &, size_t receptor_type)
virtual size_t	handles_test_event (DSSpikeEvent &, size_t receptor_type)
virtual size_t	handles_test_event (DSCurrentEvent &, size_t receptor_type)
virtual size_t	handles_test_event (GapJunctionEvent &, size_t receptor_type)
virtual size_t	handles_test_event (InstantaneousRateConnectionEvent &, size_t receptor_type)
virtual size_t	handles_test_event (DiffusionConnectionEvent &, size_t receptor_type)
virtual size_t	handles_test_event (DelayedRateConnectionEvent &, size_t receptor_type)
virtual size_t	handles_test_event (LearningSignalConnectionEvent &, size_t receptor_type)
virtual size_t	handles_test_event (SICEvent &, size_t receptor_type)
virtual void	sends_secondary_event (GapJunctionEvent &ge)
	Required to check, if source neuron may send a SecondaryEvent.
virtual void	sends_secondary_event (InstantaneousRateConnectionEvent &re)
	Required to check, if source neuron may send a SecondaryEvent.
virtual void	sends_secondary_event (DiffusionConnectionEvent &de)
	Required to check, if source neuron may send a SecondaryEvent.
virtual void	sends_secondary_event (DelayedRateConnectionEvent &re)
	Required to check, if source neuron may send a SecondaryEvent.
virtual void	sends_secondary_event (LearningSignalConnectionEvent &re)
	Required to check if source node may send a LearningSignalConnectionEvent.
virtual void	sends_secondary_event (SICEvent &sic)
	Required to check if source node may send a SICEvent.
virtual void	register_stdp_connection (double, double)
	Register a STDP connection.
virtual long	get_shift () const
	Retrieves the temporal shift of the signal.
virtual bool	is_eprop_recurrent_node () const
	Checks if the node is part of the recurrent network and thus not a readout neuron.
virtual void	handle (SpikeEvent &e)
	Handle incoming spike events.
virtual void	handle (WeightRecorderEvent &e)
	Handle incoming weight recording events.
virtual void	handle (RateEvent &e)
	Handler for rate events.
virtual void	handle (DataLoggingRequest &e)
	Handler for universal data logging request.
virtual void	handle (DataLoggingReply &e)
	Handler for universal data logging request.
virtual void	handle (CurrentEvent &e)
	Handler for current events.
virtual void	handle (ConductanceEvent &e)
	Handler for conductance events.
virtual void	handle (DoubleDataEvent &e)
	Handler for DoubleData events.
virtual void	handle (GapJunctionEvent &e)
	Handler for gap junction events.
virtual void	handle (InstantaneousRateConnectionEvent &e)
	Handler for rate neuron events.
virtual void	handle (DiffusionConnectionEvent &e)
	Handler for rate neuron events.
virtual void	handle (DelayedRateConnectionEvent &e)
	Handler for delay rate neuron events.
virtual void	handle (LearningSignalConnectionEvent &e)
	Handler for learning signal connection events.
virtual void	handle (SICEvent &e)
	Handler for slow inward current events (SICEvents).
virtual double	get_Ca_minus () const
	Return the Ca_minus value at time Ca_t which corresponds to the time of the last update in Calcium concentration which is performed each time a Node spikes.
virtual double	get_synaptic_elements (std::string) const
	Get the number of synaptic element for the current Node at Ca_t which corresponds to the time of the last spike.
virtual int	get_synaptic_elements_vacant (std::string) const
	Get the number of vacant synaptic element for the current Node Return 0 if not overridden.
virtual int	get_synaptic_elements_connected (std::string) const
	Get the number of connected synaptic element for the current Node.
virtual std::map< std::string, double >	get_synaptic_elements () const
	Get the number of all synaptic elements for the current Node at time t.
virtual void	update_synaptic_elements (double)
	Triggers the update of all SynapticElements stored in the synaptic_element_map_.
virtual void	decay_synaptic_elements_vacant ()
	Is used to reduce the number of synaptic elements in the node through time.
virtual void	connect_synaptic_element (std::string, int)
	Is used to update the number of connected synaptic elements (SynapticElement::z_connected_) when a synapse is formed or deleted.
virtual double	get_K_value (double t)
	return the Kminus value at t (in ms).
virtual double	get_LTD_value (double t)
virtual void	get_K_values (double t, double &Kminus, double &nearest_neighbor_Kminus, double &Kminus_triplet)
	write the Kminus, nearest_neighbor_Kminus, and Kminus_triplet values at t (in ms) to the provided locations.
virtual void	get_history (double t1, double t2, std::deque< histentry >::iterator *start, std::deque< histentry >::iterator *finish)
	return the spike history for (t1,t2].
virtual void	get_LTP_history (double t1, double t2, std::deque< histentry_extended >::iterator *start, std::deque< histentry_extended >::iterator *finish)
virtual void	get_urbanczik_history (double t1, double t2, std::deque< histentry_extended >::iterator *start, std::deque< histentry_extended >::iterator *finish, int)
virtual double	get_C_m (int comp)
virtual double	get_g_L (int comp)
virtual double	get_tau_L (int comp)
virtual double	get_tau_s (int comp)
virtual double	get_tau_syn_ex (int comp)
virtual double	get_tau_syn_in (int comp)
virtual void	compute_gradient (const long t_spike, const long t_spike_previous, double &z_previous_buffer, double &z_bar, double &e_bar, double &e_bar_reg, double &epsilon, double &weight, const CommonSynapseProperties &cp, WeightOptimizer *optimizer, bool is_flush_event, bool previous_was_flush_event, double &sum_grad, long &remaining_steps_until_cutoff, long &decay_steps)
	Compute gradient change for eprop synapses.
virtual double	compute_gradient (std::vector< long > &presyn_isis, const long t_previous_update, const long t_previous_trigger_spike, const double kappa, const bool average_gradient)
	Compute gradient change for eprop synapses.
virtual void	event_hook (DSSpikeEvent &)
	Modify Event object parameters during event delivery.
virtual void	event_hook (DSCurrentEvent &)
void	set_thread (size_t)
	Store the number of the thread to which the node is assigned.
size_t	get_thread () const
	Retrieve the number of the thread to which the node is assigned.
void	set_vp (size_t)
	Store the number of the virtual process to which the node is assigned.
size_t	get_vp () const
	Retrieve the number of the virtual process to which the node is assigned.
void	set_model_id (int)
	Set the model id.
void	set_initialized ()
	Execute post-initialization actions in node models.
virtual SignalType	sends_signal () const
virtual SignalType	receives_signal () const
Dictionary	get_status_base ()
	Return a dictionary with the node's properties.
void	set_status_base (const Dictionary &)
	Set status dictionary of a node.
bool	is_model_prototype () const
	Returns true if node is model prototype.
void	set_thread_lid (const size_t)
	set thread local index
size_t	get_thread_lid () const
	get thread local index
virtual void	set_local_device_id (const size_t lsdid)
	Sets the local device id.
virtual size_t	get_local_device_id () const
	Gets the local device id.
void	set_tmp_nc_index (size_t index)
	Set index in node collection; required by ThirdOutBuilder.
size_t	get_tmp_nc_index ()
	Return and invalidate index in node collection; required by ThirdOutBuilder.
Public Member Functions inherited from nest::FlushEventMechanism
	FlushEventMechanism ()
	Default constructor.
	FlushEventMechanism (const FlushEventMechanism &n)
	Copy constructor.
virtual	~FlushEventMechanism ()=default
	Destructor.
void	set_last_event_time (const long last_event_time)
	Sets the time the neuron last sent an event (spike or flush event).
long	get_last_event_time () const
	Gets the last time the neuron sent an event (spike or flush event).
void	pre_run_hook ()
	Re-calculates dependent parameters.
bool	flush_event_is_due (const long current_time) const
	Checks if a flush event is due at the current time.
void	get_status (Dictionary &d) const
	Retrieves parameters and adds them to the status dictionary.
void	set_status (const Dictionary &d, Node *node, const bool check_eprop_constraint=false)
	Sets and validates parameters from the status dictionary.
Public Member Functions inherited from nest::IgnoreAndSpikeMechanism
	IgnoreAndSpikeMechanism ()
	Default constructor.
	IgnoreAndSpikeMechanism (const IgnoreAndSpikeMechanism &n)
	Copy constructor.
virtual	~IgnoreAndSpikeMechanism ()=default
	Virtual destructor.
bool	spike_event_is_due (bool emit_dynamic_spike)
	Updates spike schedule and returns whether a spike should be emitted.
void	get_status (Dictionary &d) const
	Retrieves parameters and adds them to the status dictionary.
void	set_status (const Dictionary &d, Node *node)
	Sets and validates parameters from the status dictionary.

Public Attributes
double	firing_rate_reg_
	Firing rate regularization.
double	f_av_
	Average firing rate.
Public Attributes inherited from nest::Node
DeprecationWarning	deprecation_warning
	Member of DeprecationWarning class to be used by models if parameters are deprecated.

Protected Member Functions
long	model_dependent_history_shift_ () const override
	Returns the correct shift for history depending on whether it is a normal or a bsshslm_2020 model.
Protected Member Functions inherited from nest::Node
virtual void	init_state_ ()
	Configure state variables depending on runtime information.
virtual void	init_buffers_ ()
	Configure persistent internal data structures.
virtual void	set_initialized_ ()
Model &	get_model_ () const
void	set_frozen_ (bool frozen)
	Mark node as frozen.
template<typename ConcreteNode >
const ConcreteNode &	downcast (const Node &)
	Auxiliary function to downcast a Node to a concrete class derived from Node.

Protected Attributes
surrogate_gradient_function	compute_surrogate_gradient_
	Pointer to member function selected for computing the surrogate gradient.
Protected Attributes inherited from nest::EpropArchivingNode< HistEntryEpropRecurrent >
size_t	eprop_indegree_
	Number of incoming eprop synapses.
double	eprop_isi_trace_cutoff_
	Time interval from the previous spike until the cutoff of e-prop update computation between two spikes (ms).
std::vector< HistEntryEpropUpdate >	update_history_
	History of updates still needed by at least one synapse.
std::vector< HistEntryEpropRecurrent >	eprop_history_
	History of dynamic variables needed for e-prop plasticity.
const long	offset_gen_
	Offset since generator signals start from time step 1.
const long	delay_in_rec_
	Transmission delay from input to recurrent neurons.
const long	delay_rec_out_
	Transmission delay from recurrent to output neurons.
const long	delay_out_norm_
	Transmission delay between output neurons for normalization.
const long	delay_out_rec_
	Transmission delay from output neurons to recurrent neurons.
Protected Attributes inherited from nest::FlushEventMechanism
double	flush_event_send_interval_
	Interval since previous event after which a flush event is sent (ms).
long	flush_event_send_interval_steps_
	Interval since previous event after which a flush event is sent (steps).
long	last_event_time_
	Time of last spike or flush event (steps).

Private Attributes
size_t	n_spikes_
	Count of the emitted spikes for the firing rate regularization.
std::vector< HistEntryEpropFiringRateReg >	firing_rate_reg_history_
	History of the firing rate regularization.

Static Private Attributes
static std::map< std::string, surrogate_gradient_function >	surrogate_gradient_funcs_
	Maps provided names of surrogate gradients to corresponding pointers to member functions.

Detailed Description

template<bool hist_shift_required>
class nest::EpropArchivingNodeRecurrent< hist_shift_required >

Class implementing an intermediate archiving node model for recurrent node models supporting e-prop plasticity.

Member Typedef Documentation

surrogate_gradient_function

template<bool hist_shift_required>

typedef double( EpropArchivingNodeRecurrent::* nest::EpropArchivingNodeRecurrent< hist_shift_required >::surrogate_gradient_function) (double, double, double, double, double)

Defines the pointer-to-member function type for the surrogate gradient function.

Note

The typename is surrogate_gradient_function. All parentheses in the expression are required.

Constructor & Destructor Documentation

EpropArchivingNodeRecurrent() [1/2]

template<bool hist_shift_required>

nest::EpropArchivingNodeRecurrent< hist_shift_required >::EpropArchivingNodeRecurrent

(

)

Constructs a new EpropArchivingNodeRecurrent object.

EpropArchivingNodeRecurrent() [2/2]

template<bool hist_shift_required>

nest::EpropArchivingNodeRecurrent< hist_shift_required >::EpropArchivingNodeRecurrent

(

const EpropArchivingNodeRecurrent< hist_shift_required > &

other

)

Constructs an EpropArchivingNodeRecurrent object by copying another EpropArchivingNodeRecurrent object.

Parameters

other

The EpropArchivingNodeRecurrent object to copy.

Member Function Documentation

append_new_eprop_history_entry()

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::append_new_eprop_history_entry

(

const long

time_step

)

Creates an entry for the specified time step at the end of the eprop history.

Parameters

time_step

The time step.

compute_arctan_derivative_surrogate_gradient()

template<bool hist_shift_required>

double nest::EpropArchivingNodeRecurrent< hist_shift_required >::compute_arctan_derivative_surrogate_gradient	(	const double	r,
		const double	v_m,
		const double	v_th,
		const double	height,
		const double	width
	)

Computes the derivative of an arctan surrogate function as the surrogate gradient.

This function returns a surrogate gradient of the hard threshold function around v_th. The derivative of an arctan surrogate function is used as surrogate gradient, for example, in Fang et al. (2021). This parametrization absorbs the pi factors from Fang et al. (2021) into the height and width parameters.

Parameters

r	The number of remaining refractory steps. If r > 0, the returned surrogate gradient is zero.
v_m	The membrane voltage.
v_th	The spike threshold voltage. For adaptive neurons, this is the adaptive spike threshold voltage.
height	The positive height scaling of the surrogate gradient. This is the maximum value at v_m = v_th.
width	The positive width scaling of the surrogate gradient. Larger width gives a wider gradient.

Returns

The surrogate gradient with respect to the membrane voltage.

compute_exponential_surrogate_gradient()

template<bool hist_shift_required>

double nest::EpropArchivingNodeRecurrent< hist_shift_required >::compute_exponential_surrogate_gradient	(	const double	r,
		const double	v_m,
		const double	v_th,
		const double	height,
		const double	width
	)

Computes an exponential function as the surrogate gradient.

This function returns a surrogate gradient of the hard threshold function around v_th. The exponential surrogate gradient is used, for example, in Shrestha and Orchard (2018).

Parameters

r	The number of remaining refractory steps. If r > 0, the returned surrogate gradient is zero.
v_m	The membrane voltage.
v_th	The spike threshold voltage. For adaptive neurons, this is the adaptive spike threshold voltage.
height	The positive height scaling of the surrogate gradient. This is the maximum value at v_m = v_th.
width	The positive width scaling of the surrogate gradient. Larger width gives a wider gradient.

Returns

The surrogate gradient with respect to the membrane voltage.

compute_fast_sigmoid_derivative_surrogate_gradient()

template<bool hist_shift_required>

double nest::EpropArchivingNodeRecurrent< hist_shift_required >::compute_fast_sigmoid_derivative_surrogate_gradient	(	const double	r,
		const double	v_m,
		const double	v_th,
		const double	height,
		const double	width
	)

Computes the derivative of a fast-sigmoid surrogate function as the surrogate gradient.

This function returns a surrogate gradient of the hard threshold function around v_th. The derivative of a fast-sigmoid function is used as surrogate gradient, for example, in Zenke and Ganguli (2018).

Parameters

r	The number of remaining refractory steps. If r > 0, the returned surrogate gradient is zero.
v_m	The membrane voltage.
v_th	The spike threshold voltage. For adaptive neurons, this is the adaptive spike threshold voltage.
height	The positive height scaling of the surrogate gradient. This is the maximum value at v_m = v_th.
width	The positive width scaling of the surrogate gradient. Larger width gives a wider gradient.

Returns

The surrogate gradient with respect to the membrane voltage.

compute_piecewise_linear_surrogate_gradient()

template<bool hist_shift_required>

double nest::EpropArchivingNodeRecurrent< hist_shift_required >::compute_piecewise_linear_surrogate_gradient	(	const double	r,
		const double	v_m,
		const double	v_th,
		const double	height,
		const double	width
	)

Computes a piecewise-linear function as the surrogate gradient.

This function returns a surrogate gradient of the hard threshold function around v_th. The piecewise-linear surrogate gradient is used, for example, in Bellec et al. (2020).

Parameters

r	The number of remaining refractory steps. If r > 0, the returned surrogate gradient is zero.
v_m	The membrane voltage.
v_th	The spike threshold voltage. For adaptive neurons, this is the adaptive spike threshold voltage.
height	The positive height scaling of the surrogate gradient. This is the maximum value at v_m = v_th.
width	The positive width scaling of the surrogate gradient. Larger width gives a wider gradient.

Returns

The surrogate gradient with respect to the membrane voltage.

count_spike()

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::count_spike ( )

inline

Counts an emitted spike for the firing rate regularization.

erase_used_firing_rate_reg_history()

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::erase_used_firing_rate_reg_history

(

)

Erases the history of the used firing rate regularization history.

Erases parts of the firing rate regularization history for which the access counter in the update history has decreased to zero since no synapse needs them any longer.

find_surrogate_gradient()

template<bool hist_shift_required>

EpropArchivingNodeRecurrent< hist_shift_required >::surrogate_gradient_function nest::EpropArchivingNodeRecurrent< hist_shift_required >::find_surrogate_gradient

(

const std::string &

surrogate_gradient_function_name

)

Validates and finds surrogate gradient function based on the specified name.

Parameters

surrogate_gradient_function_name

The name of the surrogate gradient function.

Returns

The selected surrogate gradient function.

Referenced by nest::eprop_iaf::Parameters_::set(), nest::eprop_iaf_adapt::Parameters_::set(), nest::eprop_iaf_adapt_bsshslm_2020::Parameters_::set(), nest::eprop_iaf_bsshslm_2020::Parameters_::set(), nest::eprop_iaf_psc_delta::Parameters_::set(), and nest::eprop_iaf_psc_delta_adapt::Parameters_::set().

Here is the caller graph for this function:

get_firing_rate_reg_history()

template<bool hist_shift_required>

double nest::EpropArchivingNodeRecurrent< hist_shift_required >::get_firing_rate_reg_history

(

const long

time_step

)

Retrieves the firing rate regularization at the specified time step from the firing rate regularization history.

Parameters

time_step

The time step.

Returns

The firing rate regularization at the specified time step.

get_learning_signal_from_history()

template<bool hist_shift_required>

double nest::EpropArchivingNodeRecurrent< hist_shift_required >::get_learning_signal_from_history

(

const long

time_step

)

Retrieves the learning signal from the eprop history at the specified time step.

Parameters

time_step

The time step.

Returns

The learning signal at the specified time step or zero if time step is not in the history.

get_status()

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::get_status ( Dictionary &  ) const

inlineoverridevirtual

Export properties of the node by setting entries in the status dictionary.

Parameters

d	Dictionary.

Implements nest::Node.

References nest::names::eprop_isi_trace_cutoff(), nest::FlushEventMechanism::get_status(), and nest::IgnoreAndSpikeMechanism::get_status().

Referenced by nest::eprop_iaf::get_status(), nest::eprop_iaf_adapt::get_status(), nest::eprop_iaf_adapt_bsshslm_2020::get_status(), nest::eprop_iaf_bsshslm_2020::get_status(), nest::eprop_iaf_psc_delta::get_status(), and nest::eprop_iaf_psc_delta_adapt::get_status().

Here is the call graph for this function:

Here is the caller graph for this function:

model_dependent_history_shift_()

template<bool hist_shift_required>

long nest::EpropArchivingNodeRecurrent< hist_shift_required >::model_dependent_history_shift_ ( ) const

overrideprotectedvirtual

Returns the correct shift for history depending on whether it is a normal or a bsshslm_2020 model.

Returns

The history shift (steps).

Implements nest::EpropArchivingNode< HistEntryEpropRecurrent >.

reset_spike_count()

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::reset_spike_count ( )

inline

Resets the spike count for the firing rate regularization.

set_status()

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::set_status ( const Dictionary &  )

inlineoverridevirtual

Change properties of the node according to the entries in the dictionary.

Parameters

d	Dictionary with named parameter settings.

Implements nest::Node.

References nest::names::eprop_isi_trace_cutoff(), nest::IgnoreAndSpikeMechanism::set_status(), nest::FlushEventMechanism::set_status(), and nest::update_value_param().

Referenced by nest::eprop_iaf::set_status(), nest::eprop_iaf_adapt::set_status(), nest::eprop_iaf_adapt_bsshslm_2020::set_status(), nest::eprop_iaf_bsshslm_2020::set_status(), nest::eprop_iaf_psc_delta::set_status(), and nest::eprop_iaf_psc_delta_adapt::set_status().

Here is the call graph for this function:

Here is the caller graph for this function:

write_firing_rate_reg_to_history() [1/2]

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::write_firing_rate_reg_to_history	(	const long	t_current_update,
		const double	f_target,
		const double	c_reg
	)

Calculates the firing rate regularization for the current update and writes it to a new entry in the firing rate regularization history.

Parameters

t_current_update	The current update time.
f_target	The target firing rate.
c_reg	The firing rate regularization coefficient.

write_firing_rate_reg_to_history() [2/2]

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::write_firing_rate_reg_to_history	(	const long	time_step,
		const double	z,
		const double	f_target,
		const double	kappa_reg,
		const double	c_reg
	)

Calculates the current firing rate regularization and writes it to the eprop history at the specified time step.

Parameters

time_step	The time step.
z	The spike state variable.
f_target	The target firing rate.
kappa_reg	The low-pass filter of the firing rate regularization.
c_reg	The firing rate regularization coefficient.

write_learning_signal_to_history()

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::write_learning_signal_to_history	(	const long	time_step,
		const double	learning_signal
	)

Writes the learning signal to the eprop history entry at the specifed time step.

Updates the learning signal in the eprop history entry of the specified time step by writing the value of the incoming learning signal to the history or adding it to the existing value in case of multiple readout neurons.

Parameters

time_step	The time step.
learning_signal	The learning signal.

write_surrogate_gradient_to_history()

template<bool hist_shift_required>

void nest::EpropArchivingNodeRecurrent< hist_shift_required >::write_surrogate_gradient_to_history	(	const long	time_step,
		const double	surrogate_gradient
	)

Writes the surrogate gradient to the eprop history entry at the specified time step.

Parameters

time_step	The time step.
surrogate_gradient	The surrogate gradient.

Member Data Documentation

compute_surrogate_gradient_

template<bool hist_shift_required>

surrogate_gradient_function nest::EpropArchivingNodeRecurrent< hist_shift_required >::compute_surrogate_gradient_

protected

Initial value:

=
    &EpropArchivingNodeRecurrent::compute_piecewise_linear_surrogate_gradient

Pointer to member function selected for computing the surrogate gradient.

Referenced by nest::eprop_iaf::Parameters_::set(), nest::eprop_iaf_adapt::Parameters_::set(), nest::eprop_iaf_adapt_bsshslm_2020::Parameters_::set(), nest::eprop_iaf_bsshslm_2020::Parameters_::set(), nest::eprop_iaf_psc_delta::Parameters_::set(), and nest::eprop_iaf_psc_delta_adapt::Parameters_::set().

f_av_

template<bool hist_shift_required>

double nest::EpropArchivingNodeRecurrent< hist_shift_required >::f_av_

Average firing rate.

firing_rate_reg_

template<bool hist_shift_required>

double nest::EpropArchivingNodeRecurrent< hist_shift_required >::firing_rate_reg_

Firing rate regularization.

firing_rate_reg_history_

template<bool hist_shift_required>

std::vector< HistEntryEpropFiringRateReg > nest::EpropArchivingNodeRecurrent< hist_shift_required >::firing_rate_reg_history_

private

History of the firing rate regularization.

n_spikes_

template<bool hist_shift_required>

size_t nest::EpropArchivingNodeRecurrent< hist_shift_required >::n_spikes_

private

Count of the emitted spikes for the firing rate regularization.

surrogate_gradient_funcs_

template<bool hist_shift_required>

std::map< std::string, typename EpropArchivingNodeRecurrent< hist_shift_required >::surrogate_gradient_function > nest::EpropArchivingNodeRecurrent< hist_shift_required >::surrogate_gradient_funcs_

staticprivate

Initial value:

= {
    { "piecewise_linear",
      &EpropArchivingNodeRecurrent< hist_shift_required >::compute_piecewise_linear_surrogate_gradient },
    { "exponential", &EpropArchivingNodeRecurrent< hist_shift_required >::compute_exponential_surrogate_gradient },
    { "fast_sigmoid_derivative",
      &EpropArchivingNodeRecurrent< hist_shift_required >::compute_fast_sigmoid_derivative_surrogate_gradient },
    { "arctan_derivative",
      &EpropArchivingNodeRecurrent< hist_shift_required >::compute_arctan_derivative_surrogate_gradient }
  }

Maps provided names of surrogate gradients to corresponding pointers to member functions.

Todo:

In the long run, this map should be handled by a manager with proper registration functions, so that external modules can add their own gradient functions.

---

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

• eprop_archiving_node_recurrent.h
• eprop_archiving_node_recurrent_impl.h