Skip to content

v2.4.0
Compare
Choose a tag to compare
@jougs jougs released this 30 Sep 19:19
· 19242 commits to master since this release
v2.4.0
dfa61d4

License

NEST 2.4.0 is distributed under the GNU General Public License version 2 (or later) and contains many new features, neuron models and general improvements. The most important ones are:

  • A Python 3.0 compatible re-implementation of the Python interface.
  • A new framework for setting up connections and their parameters.
  • A new spike detection mode with greatly improved I/O performance.
  • More flexible framework for working with random distributions.
  • Support for connectivity-generating libraries through libneurosim.

All users are encouraged to upgrade and adapt their simulation scripts to the changes in the user interface at this point in time to benefit from the bug fixes and improvements.

Re-implementation of PyNEST

The Python interface to NEST (PyNEST) has been re-implemented from scratch. The new implementation (CyNEST) is based on Cython and provides the following improvements over the previous version:

  • Support for Python 2.6, 2.7, and 3.x
  • Installation now works properly on all platforms
  • Better extensibility and maintainability

See Zaytsev & Morrison (2014), doi:10.3389/fninf.2014.00023 for details of the new interface.

Minor changes to PyNEST are:

  • PyNEST's visualization.plot_network() can now create PNGs or PDFs
  • Importing SciPy after NEST does not cause segfaults anymore

Please note that CyNEST requires Cython version 0.19.2 or higher to regenerate the source files (only for developers). This can be installed using pip install --user --upgrade cython.

New routines for setting up connectivity

The Connect function of NEST has been completely re-written in order to support a more flexible setup of connectivity. In particular this means that connectivity is now specified using a rule ("all-to-all", "one-to-one", "fixed-indegree", "fixed-outdegree", "fixed-total-number", or "pairwise-bernoulli") and a dictionary with synapse parameters. All connection parameters can now be randomized already during connection setup by specifying the random distribution.

In addition to more flexibility, the new framework is faster due to a massive clean-up of the code and OpenMP parallelization throughout.

Please see http://www.nest-simulator.org/connection_management for more information about the new connection routines and how to transition your simulation scripts from the old functions to the new framework.

If you are using SLI, all your scripts will be working without changes. PyNEST's Connect function was renamed to OneToOneConnect. Connect now uses the new semantics as explained in the documentation.

The new Connect function is provided as a technology preview. This means that while the basic user interface will stay, the functionality and performance will be further extended in future releases.

Global spike detector

In parallel simulations on very large machines without local disks, saving recorded spikes often leads to performance problems, because each spike detector will write out one file per virtual process. This problem is now ameliorated by a new spike detection mode, which can be activated by the users of such machines. In the new mode, all spike detectors are allocated on a distinct set of processes.

See http://www.nest.simulator/global_spike_detector for details on the new recording mode and how to activate it.

Random number generators and distributions

In the process of implementing the new connection routines, we found several performance problems in NEST's librandom and inconsistencies with PyNN ([1]). To solve these problems, we changed several aspects of librandom.

The random deviate generators in NEST have been extended and modified to support random initialisation of synapse parameters for the new connection routine, and to achieve greater similarity between NEST and PyNN. For most users, these changes only add new features. Scripts using uniformint or normal_clipped need to be adapted as explained in http://www.nest-simulator.org/random_numbers.

To achieve consistent results for both global and local spike detector mode, the seed of the global RNG was changed from n_vp+1 to 0. To allow the comparison of new results with those obtained with earlier versions of NEST, this seed has to be set to n_vp+1 manually. See http://www.nest.simulator/global_spike_detector for details.

If NEST is compiled with support for GSL, gsl_rng_knuth2002 is now used instead of knuthlfg as per-thread and global random number generator (RNG) to improve performance. Both RNGs create identical sequences, so new simulation results are still comparable to old ones.

Support for connectivity-generating libraries

In NEST 2.4, the direct implementation of the ConnectionGenerator interface has been replaced in favor of support for libneurosim (http://software.incf.org/software/libneurosim). This allows to couple all connectivity-generating libraries (e.g. the Connection-set algebra; http://software.incf.org/software/csa) with NEST, which support libneurosim.

See Djurfeldt et al. (2014), doi:10.3389/fninf.2014.00043 for details about the new interface.

New models and model improvements

  • stdp_facetshw_synapse mimics the restrictions of the neuromorphic HMF developed in the context of FACETS and BrainScaleS.
  • aeif_cond_alpha_RK5, which is independent of the GSL by using a custom version of the RK5 solver.
  • aeif_cond_alpha_multisynapse is a variant of the exponential integrate-and-fire model with multiple synaptic time constants.
  • quantal_std_synapse is a probabilistic synapse model with short term plasticity (Fuhrmann et al. 2002, doi:10.1152/jn.00258.2001).
  • iaf_chs_2007 is a spike-response model (Carandini et al. 2007, doi:10.1167/7.14.20).
  • iaf_chxk_2008 is a conductance based leaky integrate-and-fire neuron model (Casti et al. 2008, doi:10.1007/s10827-007-0053-7).
  • Weighted excitatory and inhibitory input spikes are recordable in iaf_psc_exp and iaf_psc_delta.
  • Excitatory and inhibitory synaptic currents are recordable in iaf_psc_exp and iaf_psc_alpha.
  • Recovery variable U_m is recordable in izhikevich.
  • spike_detector now throws an error for spike times of 0.

Improved quality and documentation

Compared to v2.2.2, we have increased the number of unit tests from 270 to 418. This improvement guarantees the continued quality of NEST on the computer of the user and allows to find and fix problems quickly. To run the testsuite after installation, run make installcheck from the build directory.

Many examples have been updated and extended to demonsrate the usage of NEST. In addition, we have added the full microcircuituit model by Potjans & Diesmann (2014), doi:10.1093/cercor/bhs358.

In order to prevent memory leaks by forgotten arguments on the SLI stack, PyNEST now can do stack checking and runs its own testsuite in this mode.

New SLI functions and improvements

  • Added an eval function to execute SLI code in strings.
  • SetMaxBuffered which sets the max buffered parameter of a MUSIC input port.
  • eq_dv and eq_iv to test double and int vector equality.
  • cva applied to an array now leaves array elements as they are if they can not be converted to an array.
  • round now converts to double if applied on an integer.
  • cvi is identity operation for integers.
  • Take now also works for strings.

Deprecated functionality

The following SLI/PyNEST functions are obsolete will be removed in the next version:

  • The functionality of the functions (Random)DivergentConnect and (Random)ConvergentConnect is now integrated into Connect. See http://www.nest-simulator.org/connection_management for documentation on how to convert your scripts to the new syntax.
  • The old connect framework in PyNEST (OneToOneConnect) will be removed in favor of the new connect framework.
  • FindConnections is superseded by GetConnections.
  • Support for plain POSIX threads will be removed in favor of OpenMP.

Contributors since v2.2.2

  • Claudia Bachmann
  • Hannah Bos
  • Ekatarina Brocke
  • David Dahmen
  • Moritz Deger
  • Markus Diesmann
  • Mikael Djurfeldt
  • Håkon Enger
  • Jochen Martin Eppler
  • Marc-Oliver Gewaltig
  • Moritz Helias
  • Tammo Ippen
  • Jakob J. Jordan
  • Susanne Kunkel
  • Abigail Morrison
  • Mikael Naveau
  • Daniel Peppicelli
  • Alexander Peyser
  • Thomas Pfeil
  • Hans Ekkehard Plesser
  • Wolfram Schenck
  • Maximilian Schmidt
  • Jannis Schuecker
  • Sacha van Albada
  • Yury V. Zaytsev

WARNING!

The Topology Module of this version of NEST contains a bug which leads to the creation of too few connections (N/num_threads instead of N) when using the divergent connection_type in ConnectLayers in a multi-threaded mode. Please update your installation to at least v2.4.1!

Assets 3
Loading