Generalize recurrent-to-readout and readout-to-recurrent delays

models/eprop_iaf.cpp

@@ -85,6 +85,9 @@ eprop_iaf::Parameters_::Parameters_()
   , kappa_( 0.97 )
   , kappa_reg_( 0.97 )
   , eprop_isi_trace_cutoff_( 1000.0 )
+  , delay_rec_out_( 1 )
+  , delay_out_rec_( 1 )
+  , delay_total_( 1 )
 {
 }
 
@@ -131,6 +134,8 @@ eprop_iaf::Parameters_::get( DictionaryDatum& d ) const
   def< double >( d, names::kappa, kappa_ );
   def< double >( d, names::kappa_reg, kappa_reg_ );
   def< double >( d, names::eprop_isi_trace_cutoff, eprop_isi_trace_cutoff_ );
+  def< double >( d, names::delay_rec_out, Time( Time::step( delay_rec_out_ ) ).get_ms() );
+  def< double >( d, names::delay_out_rec, Time( Time::step( delay_out_rec_ ) ).get_ms() );
 }
 
 double
@@ -169,6 +174,14 @@ eprop_iaf::Parameters_::set( const DictionaryDatum& d, Node* node )
   updateValueParam< double >( d, names::kappa_reg, kappa_reg_, node );
   updateValueParam< double >( d, names::eprop_isi_trace_cutoff, eprop_isi_trace_cutoff_, node );
 
+  double delay_rec_out_ms = Time( Time::step( delay_rec_out_ ) ).get_ms();
+  updateValueParam< double >( d, names::delay_rec_out, delay_rec_out_ms, node );
+  delay_rec_out_ = Time( Time::ms( delay_rec_out_ms ) ).get_steps();
+
+  double delay_out_rec_ms = Time( Time::step( delay_out_rec_ ) ).get_ms();
+  updateValueParam< double >( d, names::delay_out_rec, delay_out_rec_ms, node );
+  delay_out_rec_ = Time( Time::ms( delay_out_rec_ms ) ).get_steps();
+
   if ( C_m_ <= 0 )
   {
     throw BadProperty( "Membrane capacitance C_m > 0 required." );
@@ -214,6 +227,18 @@ eprop_iaf::Parameters_::set( const DictionaryDatum& d, Node* node )
     throw BadProperty( "Cutoff of integration of eprop trace between spikes eprop_isi_trace_cutoff ≥ 0 required." );
   }
 
+  if ( delay_rec_out_ < 1 )
+  {
+    throw BadProperty( "Connection delay from recurrent to readout neuron ≥ 1 required." );
+  }
+
+  if ( delay_out_rec_ < 1 )
+  {
+    throw BadProperty( "Connection delay from readout to recurrent neuron ≥ 1 required." );
+  }
+
+  delay_total_ = delay_rec_out_ + ( delay_out_rec_ - 1 );
+
   return delta_EL;
 }
 
@@ -278,6 +303,14 @@ eprop_iaf::pre_run_hook()
 
   V_.P_v_m_ = std::exp( -dt / P_.tau_m_ );
   V_.P_i_in_ = P_.tau_m_ / P_.C_m_ * ( 1.0 - V_.P_v_m_ );
+
+  if ( eprop_history_.empty() )
+  {
+    for ( long t = -P_.delay_total_; t < 0; ++t )
+    {
+      append_new_eprop_history_entry( t );
+    }
+  }
 }
 
 
@@ -373,7 +406,8 @@ eprop_iaf::handle( DataLoggingRequest& e )
 void
 eprop_iaf::compute_gradient( const long t_spike,
   const long t_spike_previous,
-  double& z_previous_buffer,
+  std::queue< double >& z_previous_buffer,
+  double& z_previous,
   double& z_bar,
   double& e_bar,
   double& e_bar_reg,
@@ -382,26 +416,31 @@ eprop_iaf::compute_gradient( const long t_spike,
   const CommonSynapseProperties& cp,
   WeightOptimizer* optimizer )
 {
-  double e = 0.0;                // eligibility trace
-  double z = 0.0;                // spiking variable
-  double z_current_buffer = 1.0; // buffer containing the spike that triggered the current integration
-  double psi = 0.0;              // surrogate gradient
-  double L = 0.0;                // learning signal
-  double firing_rate_reg = 0.0;  // firing rate regularization
-  double grad = 0.0;             // gradient
+  double e = 0.0;               // eligibility trace
+  double z = 0.0;               // spiking variable
+  double z_current = 1.0;       // spike state that triggered the current integration
+  double psi = 0.0;             // surrogate gradient
+  double L = 0.0;               // learning signal
+  double firing_rate_reg = 0.0; // firing rate regularization
+  double grad = 0.0;            // gradient
 
   const EpropSynapseCommonProperties& ecp = static_cast< const EpropSynapseCommonProperties& >( cp );
   const auto optimize_each_step = ( *ecp.optimizer_cp_ ).optimize_each_step_;
 
-  auto eprop_hist_it = get_eprop_history( t_spike_previous - 1 );
+  auto eprop_hist_it = get_eprop_history( t_spike_previous - P_.delay_total_ );
 
   const long t_compute_until = std::min( t_spike_previous + V_.eprop_isi_trace_cutoff_steps_, t_spike );
 
   for ( long t = t_spike_previous; t < t_compute_until; ++t, ++eprop_hist_it )
   {
-    z = z_previous_buffer;
-    z_previous_buffer = z_current_buffer;
-    z_current_buffer = 0.0;
+    if ( P_.delay_total_ > 1 )
+    {
+      update_pre_syn_buffer_multiple_entries( z, z_current, z_previous, z_previous_buffer, t_spike, t );
+    }
+    else
+    {
+      update_pre_syn_buffer_one_entry( z, z_current, z_previous, z_previous_buffer, t_spike, t );
+    }
 
     psi = eprop_hist_it->surrogate_gradient_;
     L = eprop_hist_it->learning_signal_;

models/eprop_iaf.h

@@ -390,6 +390,7 @@ class eprop_iaf : public EpropArchivingNodeRecurrent< false >
 
   void compute_gradient( const long,
     const long,
+    std::queue< double >&,
     double&,
     double&,
     double&,
@@ -402,6 +403,9 @@ class eprop_iaf : public EpropArchivingNodeRecurrent< false >
   long get_shift() const override;
   bool is_eprop_recurrent_node() const override;
   long get_eprop_isi_trace_cutoff() const override;
+  long get_delay_total() const override;
+  long get_delay_recurrent_to_readout() const override;
+  long get_delay_readout_to_recurrent() const override;
 
   //! Map for storing a static set of recordables.
   friend class RecordablesMap< eprop_iaf >;
@@ -458,6 +462,15 @@ class eprop_iaf : public EpropArchivingNodeRecurrent< false >
     //! Time interval from the previous spike until the cutoff of e-prop update integration between two spikes (ms).
     double eprop_isi_trace_cutoff_;
 
+    //! Connection delay from recurrent to readout neuron.
+    long delay_rec_out_;
+
+    //! Connection delay from readout to recurrent neuron.
+    long delay_out_rec_;
+
+    //! Sum of connection delays from recurrent to readout neuron and from readout to recurrent neuron.
+    long delay_total_;
+
     //! Default constructor.
     Parameters_();
 
@@ -594,10 +607,35 @@ eprop_iaf::get_eprop_isi_trace_cutoff() const
   return V_.eprop_isi_trace_cutoff_steps_;
 }
 
+inline long
+eprop_iaf::get_delay_total() const
+{
+  return P_.delay_total_;
+}
+
+inline long
+eprop_iaf::get_delay_recurrent_to_readout() const
+{
+  return P_.delay_rec_out_;
+}
+
+inline long
+eprop_iaf::get_delay_readout_to_recurrent() const
+{
+  return P_.delay_out_rec_;
+}
+
 inline size_t
 eprop_iaf::send_test_event( Node& target, size_t receptor_type, synindex, bool )
 {
   SpikeEvent e;
+
+  // To perform a consistency check on the delay parameter d_out_rec between recurrent
+  // neurons and readout neurons, the recurrent neurons send a test event with a delay
+  // specified by d_rec_out. Upon receiving the test event from the recurrent neuron,
+  // the readout neuron checks if the delay with which the event was received matches
+  // its own specified delay parameter d_rec_out.
+  e.set_delay_steps( P_.delay_rec_out_ );
   e.set_sender( *this );
   return target.handles_test_event( e, receptor_type );
 }

models/eprop_iaf_adapt.cpp

@@ -89,6 +89,9 @@ eprop_iaf_adapt::Parameters_::Parameters_()
   , kappa_( 0.97 )
   , kappa_reg_( 0.97 )
   , eprop_isi_trace_cutoff_( 1000.0 )
+  , delay_rec_out_( 1 )
+  , delay_out_rec_( 1 )
+  , delay_total_( 1 )
 {
 }
 
@@ -139,6 +142,8 @@ eprop_iaf_adapt::Parameters_::get( DictionaryDatum& d ) const
   def< double >( d, names::kappa, kappa_ );
   def< double >( d, names::kappa_reg, kappa_reg_ );
   def< double >( d, names::eprop_isi_trace_cutoff, eprop_isi_trace_cutoff_ );
+  def< double >( d, names::delay_rec_out, Time( Time::step( delay_rec_out_ ) ).get_ms() );
+  def< double >( d, names::delay_out_rec, Time( Time::step( delay_out_rec_ ) ).get_ms() );
 }
 
 double
@@ -179,6 +184,12 @@ eprop_iaf_adapt::Parameters_::set( const DictionaryDatum& d, Node* node )
   updateValueParam< double >( d, names::kappa_reg, kappa_reg_, node );
   updateValueParam< double >( d, names::eprop_isi_trace_cutoff, eprop_isi_trace_cutoff_, node );
 
+  const double delay_rec_out_ = Time::step( delay_rec_out_ );
+  updateValueParam< double >( d, names::delay_rec_out, Time( delay_rec_out_).get_ms(), node );
+
+  const double delay_out_rec_ = Time::step( delay_out_rec_ );
+  updateValueParam< double >( d, names::delay_out_rec, Time( delay_out_rec_ ).get_ms(), node );
+
   if ( adapt_beta_ < 0 )
   {
     throw BadProperty( "Threshold adaptation prefactor adapt_beta ≥ 0 required." );
@@ -234,6 +245,18 @@ eprop_iaf_adapt::Parameters_::set( const DictionaryDatum& d, Node* node )
     throw BadProperty( "Cutoff of integration of eprop trace between spikes eprop_isi_trace_cutoff ≥ 0 required." );
   }
 
+  if ( delay_rec_out_ < 1 )
+  {
+    throw BadProperty( "Connection delay from recurrent to readout neuron ≥ 1 required." );
+  }
+
+  if ( delay_out_rec_ < 1 )
+  {
+    throw BadProperty( "Connection delay from readout to recurrent neuron ≥ 1 required." );
+  }
+
+  delay_total_ = delay_rec_out_ + ( delay_out_rec_ - 1 );
+
   return delta_EL;
 }
 
@@ -313,6 +336,14 @@ eprop_iaf_adapt::pre_run_hook()
   V_.P_v_m_ = std::exp( -dt / P_.tau_m_ );
   V_.P_i_in_ = P_.tau_m_ / P_.C_m_ * ( 1.0 - V_.P_v_m_ );
   V_.P_adapt_ = std::exp( -dt / P_.adapt_tau_ );
+
+  if ( eprop_history_.empty() )
+  {
+    for ( long t = -P_.delay_total_; t < 0; ++t )
+    {
+      append_new_eprop_history_entry( t );
+    }
+  }
 }
 
 
@@ -412,7 +443,8 @@ eprop_iaf_adapt::handle( DataLoggingRequest& e )
 void
 eprop_iaf_adapt::compute_gradient( const long t_spike,
   const long t_spike_previous,
-  double& z_previous_buffer,
+  std::queue< double >& z_previous_buffer,
+  double& z_previous,
   double& z_bar,
   double& e_bar,
   double& e_bar_reg,
@@ -421,26 +453,31 @@ eprop_iaf_adapt::compute_gradient( const long t_spike,
   const CommonSynapseProperties& cp,
   WeightOptimizer* optimizer )
 {
-  double e = 0.0;                // eligibility trace
-  double z = 0.0;                // spiking variable
-  double z_current_buffer = 1.0; // buffer containing the spike that triggered the current integration
-  double psi = 0.0;              // surrogate gradient
-  double L = 0.0;                // learning signal
-  double firing_rate_reg = 0.0;  // firing rate regularization
-  double grad = 0.0;             // gradient
+  double e = 0.0;               // eligibility trace
+  double z = 0.0;               // spiking variable
+  double z_current = 1.0;       // spike state that triggered the current integration
+  double psi = 0.0;             // surrogate gradient
+  double L = 0.0;               // learning signal
+  double firing_rate_reg = 0.0; // firing rate regularization
+  double grad = 0.0;            // gradient
 
   const EpropSynapseCommonProperties& ecp = static_cast< const EpropSynapseCommonProperties& >( cp );
   const auto optimize_each_step = ( *ecp.optimizer_cp_ ).optimize_each_step_;
 
-  auto eprop_hist_it = get_eprop_history( t_spike_previous - 1 );
+  auto eprop_hist_it = get_eprop_history( t_spike_previous - P_.delay_total_ );
 
   const long t_compute_until = std::min( t_spike_previous + V_.eprop_isi_trace_cutoff_steps_, t_spike );
 
   for ( long t = t_spike_previous; t < t_compute_until; ++t, ++eprop_hist_it )
   {
-    z = z_previous_buffer;
-    z_previous_buffer = z_current_buffer;
-    z_current_buffer = 0.0;
+    if ( P_.delay_total_ > 1 )
+    {
+      update_pre_syn_buffer_multiple_entries( z, z_current, z_previous, z_previous_buffer, t_spike, t );
+    }
+    else
+    {
+      update_pre_syn_buffer_one_entry( z, z_current, z_previous, z_previous_buffer, t_spike, t );
+    }
 
     psi = eprop_hist_it->surrogate_gradient_;
     L = eprop_hist_it->learning_signal_;

models/eprop_iaf_adapt.h

@@ -358,6 +358,7 @@ class eprop_iaf_adapt : public EpropArchivingNodeRecurrent< false >
 
   void compute_gradient( const long,
     const long,
+    std::queue< double >&,
     double&,
     double&,
     double&,
@@ -370,6 +371,9 @@ class eprop_iaf_adapt : public EpropArchivingNodeRecurrent< false >
   long get_shift() const override;
   bool is_eprop_recurrent_node() const override;
   long get_eprop_isi_trace_cutoff() const override;
+  long get_delay_total() const override;
+  long get_delay_recurrent_to_readout() const override;
+  long get_delay_readout_to_recurrent() const override;
 
   //! Map for storing a static set of recordables.
   friend class RecordablesMap< eprop_iaf_adapt >;
@@ -432,6 +436,15 @@ class eprop_iaf_adapt : public EpropArchivingNodeRecurrent< false >
     //! Time interval from the previous spike until the cutoff of e-prop update integration between two spikes (ms).
     double eprop_isi_trace_cutoff_;
 
+    //! Connection delay from recurrent to readout neuron.
+    long delay_rec_out_;
+
+    //! Connection delay from readout to recurrent neuron.
+    long delay_out_rec_;
+
+    //! Sum of connection delays from recurrent to readout neuron and readout to recurrent neuron.
+    long delay_total_;
+
     //! Default constructor.
     Parameters_();
 
@@ -591,10 +604,35 @@ eprop_iaf_adapt::get_eprop_isi_trace_cutoff() const
   return V_.eprop_isi_trace_cutoff_steps_;
 }
 
+inline long
+eprop_iaf_adapt::get_delay_total() const
+{
+  return P_.delay_total_;
+}
+
+inline long
+eprop_iaf_adapt::get_delay_recurrent_to_readout() const
+{
+  return P_.delay_rec_out_;
+}
+
+inline long
+eprop_iaf_adapt::get_delay_readout_to_recurrent() const
+{
+  return P_.delay_out_rec_;
+}
+
 inline size_t
 eprop_iaf_adapt::send_test_event( Node& target, size_t receptor_type, synindex, bool )
 {
   SpikeEvent e;
+
+  // To perform a consistency check on the delay parameter d_out_rec between recurrent
+  // neurons and readout neurons, the recurrent neurons send a test event with a delay
+  // specified by d_rec_out. Upon receiving the test event from the recurrent neuron,
+  // the readout neuron checks if the delay with which the event was received matches
+  // its own specified delay parameter d_rec_out.
+  e.set_delay_steps( P_.delay_rec_out_ );
   e.set_sender( *this );
   return target.handles_test_event( e, receptor_type );
 }