Skip to content

Latest commit

 

History

History
160 lines (126 loc) · 9.77 KB

nest-2.2.x-release_notes.md

File metadata and controls

160 lines (126 loc) · 9.77 KB
Raw

Release notes for NEST 2.2.x

NEST 2.2 contains substantial improvements and many new features. The most important ones are:

  • Better speed, scaling, and memory footprint
  • Support for connection set algebra (CSA)
  • Data driven network generation

NEST 2.2.2 Release Notes

  • New models and devices:
    • iaf_psc_alpha_multisynapse
    • iaf_psc_exp_multisynapse
    • mcculloch_pitts_neuron
    • ginzburg_neuron
    • spin_detector
    • sinusoidal_gamma_generator
  • Updated models and devices:
    • smp_generator is replaced by sinusoidal_poisson_generator, which sends an '''individual''' spike train to each of its targets by default; be wary when updating your code! In order to replicate the old behavior of the smp_generator (all targets receive the same spike train), set /individual_spike_trains to false on the sinusoidal_poisson_generator model before creating a generator node.
    • iaf_psc_alpha: bugfixes
    • stdp_dopamine_synapse: bugfixes
  • PyNEST: fixes to the DataConnect() interface
  • Topology:
    • more efficient GetTargetNodes() implementation
    • lognormal distribution for parameter values
  • Various bugfixes to the kernel, SLI, build system and other areas

NEST 2.2.1 Release Notes

NEST 2.2.1 is primarily a bugfix release, resolving several bugs in the build system, PyNEST, topology module and built-in models. NEST 2.2.1 has been verified to be compatible with the latest released PyNN 0.7.5.

NEST 2.2.0 Release Notes

Better speed, scaling, and memory footprint

Many of NEST's major data structures have been re-written to improve speed, scalability and memory footprint.

The most significant changes concern the simulation kernel and the way it represents networks internally. As a result, NEST has a considerably lower memory consumption and improved scaling, in particular when using a large number of cores. The full details and theory behind these changes are published in two papers:

  • Helias et al. Front. Neuroinform. (2012) http://dx.doi.org/10.3389/fninf.2012.00026

  • Kunkel et al. Front. Neuroinform. (2012) http://dx.doi.org/10.3389/fninf.2011.00035

  • Many connect functions now use OpenMP to connect neurons in parallel.

  • OpenMP has also replaced Pthreads as a default for multi-threaded simulations, yielding better scaling.

  • SLI, the built-in simulation language interpreter of NEST, is now up to 5 times faster.

  • The topology library has been re-written, greatly improving its speed and reducing memory requirements.

Support for connection set algebra (CSA)

NEST 2.2.0 supports the Connection Set Algebra by Mikael Djurfeldt (http://dx.doi.org/10.1007/s12021-012-9146-1). The Connection Set Algebra is a powerful notation that allows to formulate complex network architectures in a concise manner. The following examples illustrates how the CSA is used to connect a random network:

import csa
import nest

# Random connectivity with connection probability 0.1
cs = csa.random(0.1)

# Create two neuron populations
pop1 = nest.LayoutNetwork("iaf_neuron", [16])
pop2 = nest.LayoutNetwork("iaf_neuron", [16])

# Connect them using the connection generator cs
nest.CGConnect (pop0, pop1, cs)

Data driven network generation

NEST 2.2.0 has a new function DataConnect which allows the efficient connection and parametrization of connections from connection data. DataConnect will efficiently create and parameterize synapses.

The new function GetConnections allows to efficiently retrieve the afferent and efferent connections of a neuron or the connections between groups of neurons. The combination of GetConnections and DataConnect allows users to retrieve, save, and restore the synaptic state of a network. This is particularly useful in models with synaptic plasticity and learning.

Topology library

The topology library supports the creation of spatially organized networks, e.g. for models of the visual system. NEST 2.2.0 supports 3-dimensional networks, where neurons are placed in a volume rather than on a sheet. The following example shows how to connect a layer to itself with a Gaussian distance-dependent probability profile:

import nest
import nest.topology as topo

# Specify layer structure and connectivity profile
layer_spec = {'columns': 30, 'rows': 30, 'extent': [3.0, 3.0], 'elements': 'iaf_neuron'} 
conn_spec = {'connection_type': 'convergent',
             'mask': {'circular': {'radius': 3.0}},
             'kernel': {'gaussian': {'p_center': 1.0, 'sigma': 0.5}},
             'weights': 1.0, 'delays': 1.0}

# Create layer and connections
layer = topo.CreateLayer(layer_spec)
topo.ConnectLayers(layer, layer, conn_spec)

# Visualize all targets of neuron at center of layer
topo.PlotTargets(topo.FindCenterElement(layer), layer)

There is also a new API to add user defined connection kernels. Please refer to the updated user manual and examples for more details.

Detailed list of changes

  • Kernel and PyNEST changes
    • Major improvements to the memory consumption and scaling properties of the simulation kernel as described in Helias et al. (2012) and Kunkel et al. (2012).
    • NEST now supports the connection generator interface, an interface allowing external modules to generate connectivity, see http://software.incf.org/software/libneurosim.
    • Support for the Connection Set Algebra (CSA) by Mikael Djurfeldt has been added, see http://software.incf.org/software/csa and http://dx.doi.org/10.3389/conf.fninf.2011.08.00085.
    • The new command GetConnections allows the fast retrieval of connections and will replace the slow and memory intensive FindConnections; FindConnections is deprecated and will be removed in future releases.
    • The new command DataConnect allows to efficiently create network connections from data, e.g. when synapse parameters are explicitly given.
    • Connection objects are now represented as Python lists or NumPy arrays; code which relies on the old connection dictionaries will have to be changed.
    • The function GetNodes has been removed, one has to explicitly use either GetLocalNodes, or GetGlobalNodes.
    • Support for node addresses has been removed and with it the functions GetAddress and GetGID.
    • Threading support for OpenMP has been added and is the default now.
    • Many connect routines and node calibration are now parallel, using OpenMP.
    • New function CGConnect for connecting neurons using connection generators (see doc/conngen.txt).
    • New function abort, which terminates NEST and all its MPI processes without deadlocks.
  • Topology module changes
    • Major rewrite which resulted in improved performance and reduced memory requirements for freely placed neurons.
    • Topology now supports 3-dimensional layers.
    • New API for adding your own kernel functions.
    • 'Nested' layer layout (subnets within subnets) is no longer supported; this was previously discouraged for performance reasons.
    • Composite layers (layers with multiple nodes per position) no longer contain subnets.
    • Semantics of GetElement have changed, in particular a list of GIDs is returned for a composite layer, where previously a single subnet GID would be returned.
    • GetPosition, Displacement and Distance now only work for nodes local to the current MPI process.
    • See the updated Topology User Manual for details.
  • SLI Interpreter improvements
    • The SLI Interpreter has been optimized for speed and memory; in particular handling and lookup of names is much faster now.
    • SLI now supports fixed size vectors of doubles and integers. The new types are called IntVector (/intvectortype) and DoubleVector (/doublevectortype).
    • NumPy arrays are automatically converted to SLI vectors to conserve memory and CPU time.
    • SLI has new functions arange, zeros and ones to easily create vectors.
    • The vector types support all common math operations.
    • New functions DictQ and SubnetQ to test the argument types for being a dictionary or a subnet, respectively.
  • Miscellaneous changes
    • The installation prefix should now be given explicitly, since installing to the default /usr/local is strongly discouraged.
    • Substantial documentation updates; a large number of broken documentation cross-references has been resolved.
    • Fixed numerical instabilities of the AdEx models (aeiaf_cond_exp and aeiaf_cond_alpha).
    • New significantly faster binomial random number generator replaced the previous implementation in librandom (#390).
    • New wrapper for the GSL binomial random deviate generator under the name gsl_binomial.
    • Support for IBM BlueGene and K supercomputers (configure with --enable-bluegene=l/p/q).
    • The command setenvironment has been removed; this functionality was broken on Mac OS X for quite some time.
    • Much improved test coverage for SLI, PyNEST and MPI in particular.
    • Many new SLI and PyNEST examples and updates for the existing ones.
    • Code quality of NEST and all examples is continuously monitored now using CI (http://dx.doi.org/10.3389/fninf.2012.00031).

Known issues

  • The simulation progress indicator does not work with OpenMP and some error messages are unreadable.
  • On multiarch systems (i.e. 64-bit Red Hat Linux) one has to manually move all PyNEST-related files from $PREFIX/lib/python2.6/site-packages to $PREFIX/lib64/python2.6/site-packages; this will be resolved in next releases. This also breaks make installcheck.
  • Known compatibility problems with PyNN are now fixed in the 0.7 svn branch, which will be released early next year.

Contributors

  • Sacha van Albada
  • Giuseppe Chindemi
  • Moritz Deger
  • Markus Diesmann
  • Mikael Djurfeldt
  • Håkon Enger
  • Jochen Martin Eppler
  • Marc-Oliver Gewaltig
  • Moritz Helias
  • Susanne Kunkel
  • Abigail Morrison
  • Eilif Muller
  • Hans Ekkehard Plesser
  • Sven Schrader
  • Tom Tetzlaff
  • Yury V. Zaytsev