Implement e-prop plasticity

build_support/generate_modelsmodule.py

@@ -27,6 +27,7 @@
 """
 
 import argparse
+import itertools
 import os
 import sys
 from pathlib import Path
@@ -109,6 +110,7 @@ def get_models_from_file(model_file):
         "public Node": "node",
         "public ClopathArchivingNode": "clopath",
         "public UrbanczikArchivingNode": "urbanczik",
+        "public EpropArchivingNode": "neuron",
         "typedef binary_neuron": "binary",
         "typedef rate_": "rate",
     }
@@ -227,9 +229,7 @@ def generate_modelsmodule():
     1. the copyright header.
     2. a list of generic NEST includes
     3. the list of includes for the models to build into NEST
-    4. some boilerplate function implementations needed to fulfill the
-       Module interface
-    5. the list of model registration lines for the models to build
+    4. the list of model registration lines for the models to build
        into NEST
 
     The code is enriched by structured C++ comments as to make
@@ -246,7 +246,16 @@ def generate_modelsmodule():
     modeldir.mkdir(parents=True, exist_ok=True)
     with open(modeldir / fname, "w") as file:
         file.write(copyright_header.replace("{{file_name}}", fname))
-        file.write('\n#include "models.h"\n\n// Generated includes\n#include "config.h"\n')
+        file.write(
+            dedent(
+                """
+            #include "models.h"
+
+            // Generated includes
+            #include "config.h"
+        """
+            )
+        )
 
         for model_type, guards_fnames in includes.items():
             file.write(f"\n// {model_type.capitalize()} models\n")

doc/htmldoc/examples/index.rst

@@ -198,6 +198,16 @@ PyNEST examples
            * :doc:`../auto_examples/evaluate_tsodyks2_synapse`
 
 
+.. grid:: 1 1 2 3
+
+    .. grid-item-card:: :doc:`../auto_examples/eprop_plasticity/index`
+           :img-top: ../static/img/pynest/eprop_supervised_classification_infrastructure.png
+
+           * :doc:`/auto_examples/eprop_plasticity/eprop_supervised_classification_evidence-accumulation`
+           * :doc:`/auto_examples/eprop_plasticity/eprop_supervised_regression_sine-waves`
+           * :doc:`/auto_examples/eprop_plasticity/eprop_supervised_regression_handwriting`
+           * :doc:`/auto_examples/eprop_plasticity/eprop_supervised_regression_infinite-loop`
+
 
 .. grid:: 1 1 2 3
 
@@ -332,6 +342,7 @@ PyNEST examples
    ../auto_examples/astrocytes/astrocyte_interaction
    ../auto_examples/astrocytes/astrocyte_small_network
    ../auto_examples/astrocytes/astrocyte_brunel
+   ../auto_examples/eprop_plasticity/index
 
 .. toctree::
    :hidden:

doc/htmldoc/whats_new/v3.7/index.rst

@@ -40,3 +40,25 @@ See examples using astrocyte models:
 See connectivity documentation:
 
 * :ref:`tripartite_connectivity`
+
+
+E-prop plasticity in NEST
+-------------------------
+
+Another new NEST feature is eligibility propagation (e-prop) [1]_, a local and
+online learning algorithm for recurrent spiking neural networks (RSNNs) that
+serves as a biologically plausible approximation to backpropagation through time
+(BPTT). It relies on eligibility traces and neuron-specific learning signals to
+compute gradients without the need for error propagation backward in time. This
+approach aligns with the brain's learning mechanisms and offers a strong
+candidate for efficient training of RSNNs in low-power neuromorphic hardware.
+
+For further information, see:
+
+* :doc:`/auto_examples/eprop_plasticity/index`
+* :doc:`/models/index_e-prop plasticity`
+
+.. [1] Bellec G, Scherr F, Subramoney F, Hajek E, Salaj D, Legenstein R,
+       Maass W (2020). A solution to the learning dilemma for recurrent
+       networks of spiking neurons. Nature Communications, 11:3625.
+       https://doi.org/10.1038/s41467-020-17236-y

libnestutil/block_vector.h

@@ -236,6 +236,14 @@ class BlockVector
    */
   void push_back( const value_type_& value );
 
+  /**
+   * @brief Move data to the end of the BlockVector.
+   * @param value Data to be moved to end of BlockVector.
+   *
+   * Moves given data to the element at the end of the BlockVector.
+   */
+  void push_back( value_type_&& value );
+
   /**
    * Erases all the elements.
    */
@@ -313,15 +321,17 @@ class BlockVector
 /////////////////////////////////////////////////////////////
 
 template < typename value_type_ >
-inline BlockVector< value_type_ >::BlockVector()
-  : blockmap_( std::vector< std::vector< value_type_ > >( 1, std::vector< value_type_ >( max_block_size ) ) )
+BlockVector< value_type_ >::BlockVector()
+  : blockmap_(
+    std::vector< std::vector< value_type_ > >( 1, std::move( std::vector< value_type_ >( max_block_size ) ) ) )
   , finish_( begin() )
 {
 }
 
 template < typename value_type_ >
-inline BlockVector< value_type_ >::BlockVector( size_t n )
-  : blockmap_( std::vector< std::vector< value_type_ > >( 1, std::vector< value_type_ >( max_block_size ) ) )
+BlockVector< value_type_ >::BlockVector( size_t n )
+  : blockmap_(
+    std::vector< std::vector< value_type_ > >( 1, std::move( std::vector< value_type_ >( max_block_size ) ) ) )
   , finish_( begin() )
 {
   size_t num_blocks_needed = std::ceil( static_cast< double >( n ) / max_block_size );
@@ -394,7 +404,7 @@ BlockVector< value_type_ >::end() const
 }
 
 template < typename value_type_ >
-inline void
+void
 BlockVector< value_type_ >::push_back( const value_type_& value )
 {
   // If this is the last element in the current block, add another block
@@ -411,7 +421,24 @@ BlockVector< value_type_ >::push_back( const value_type_& value )
 }
 
 template < typename value_type_ >
-inline void
+void
+BlockVector< value_type_ >::push_back( value_type_&& value )
+{
+  // If this is the last element in the current block, add another block
+  if ( finish_.block_it_ == finish_.current_block_end_ - 1 )
+  {
+    // Need to get the current position here, then recreate the iterator after we extend the blockmap,
+    // because after the blockmap is changed the iterator becomes invalid.
+    const auto current_block = finish_.block_vector_it_ - finish_.block_vector_->blockmap_.begin();
+    blockmap_.emplace_back( max_block_size );
+    finish_.block_vector_it_ = finish_.block_vector_->blockmap_.begin() + current_block;
+  }
+  *finish_ = std::move( value );
+  ++finish_;
+}
+
+template < typename value_type_ >
+void
 BlockVector< value_type_ >::clear()
 {
   for ( auto it = blockmap_.begin(); it != blockmap_.end(); ++it )
@@ -442,7 +469,7 @@ BlockVector< value_type_ >::size() const
 }
 
 template < typename value_type_ >
-inline typename BlockVector< value_type_ >::iterator
+typename BlockVector< value_type_ >::iterator
 BlockVector< value_type_ >::erase( const_iterator first, const_iterator last )
 {
   assert( first.block_vector_ == this );
@@ -495,7 +522,7 @@ BlockVector< value_type_ >::erase( const_iterator first, const_iterator last )
 }
 
 template < typename value_type_ >
-inline void
+void
 BlockVector< value_type_ >::print_blocks() const
 {
   std::cerr << "this: \t\t" << this << "\n";

models/CMakeLists.txt

@@ -26,6 +26,7 @@ set(models_sources
     rate_neuron_ipn.h rate_neuron_ipn_impl.h
     rate_neuron_opn.h rate_neuron_opn_impl.h
     rate_transformer_node.h rate_transformer_node_impl.h
+    weight_optimizer.h weight_optimizer.cpp
     ${MODELS_SOURCES_GENERATED}
 )