 |
|
|
|
The Need for Simulation Software
Because of the immense difficulties involved in observing neuronal
activity in an intact brain, experimental data exists mostly for
single or small numbers of neurons, or for very large
populations (106 neurons). As a consequence, scientists
increasingly employ theoretical and computational methods in order to
assess the dynamical properties of individual and populations of
neurons and to come up with testable predictions.
Computational neuroscience is, thus, a fast growing research
field, dedicated to the investigation of the nervous system with the
help of computer simulations. It is important to note, however, that
the type of neuronal network model discussed here differs considerably
from so-called artificial neural networks (ANN), which have become an
important branch of engineering and computer science.
In Computational Neuroscience, simulations are used to investigate
models of the nervous system at functional or process levels.
Consequently, a lot of effort has been put into developing appropriate
simulation tools and techniques, and a plethora of simulation
software, specialized for the single neuron or small sized networks is
available.
Recently, however, there has been growing interest in large scale
simulations, involving some 104 neurons while maintaining an
acceptable degree of biological detail. Thus, there is need for
simulation software, possibly parallel, which supports such
simulations in a flexible way.
The Neuro Simulation Tool NEST
The Neuro Simulation Tool NEST is a tool for the simulation of
large networks. It is being developed by the members of the NEST
Initiative and has been available to the public under an open source
license since August 2004. The NEST simulator is optimized for
the simulation of large networks of neurons; at present, networks of
some 100.000 neurons with around 1 billion synapses (connections) can
be simulated within reasonable time. For detailed information about
the NEST simulator, including download, please see the homepage of the
NEST Initiative.
The Computational Neuroscience Group
at UMB has been a member of the NEST Initiative since its inception in
2001. We have since made significant contributions to the following
aspects of NEST:
Parallel Simulation in NEST
The simulation of large neuronal networks requires parallel
computers, both because large amounts of memory are required and to
keep simulation times bearable. The simulation of neuronal networks
with their multitude of connections and highly variable activity poses
very different challenges when parallelizing than, e.g., matrix
algebra. The present release version of NEST shows indeed rather
disappointing scaling behavior (see figure above, red curve): When
using 8 instead of 1 processor to simulate a network of 12600 neurons
and 16 million synapses, the simulation time is reduced by a factor of
only three. Through work done at UMB, using our AlphaServer GS1280
supercomputer Caspar, we have been able to reach essentially
linear scaling (blue curve): using eight processors, the simulation
runs eight times as fast. This is excellent in view of Amdahl's Law of parallel computation, which
predicts that even linear speed-up is usually not attainable.
NEST achieves it because fast processor cache memory is exploited the
better the more processors are used. Our collaborators in Freiburg
have even demonstrated clearly superlinear scaling in a variant of the
NEST simulator (Morrison et al, to appear in Neural
Computation).
We are currently working on further improvements of parallel
performance in NEST. Our long-term goal is a simulator based on
hybrid parallelization that would be ideally suited for the
increasingly common clusters of multi-processor computers.
CoThaCo: A Comprehensive Model of the Thalamocortical
Pathway
Over the past years, several studies have demonstrated that feedback
plays an important role for visual processing in later stages of the
visual pathway, from primary visual cortex on upward. In the early visual pathway, the
picture is less clear-cut at present. If we set aside eye movements,
there is no feedback into the retina. In the dorsal lateral
geniculate nucleus (dLGN), on the other hand, only some 10% of all
synapses convey ascending input from the retina, while the vast
majority mediates descending signals from visual cortex. It appears implausible that evolution should have
retained the massive corticothalamic feedback without purpose.
Several groups have indeed demonstrated in cat
and primate that corticothalamic feedback
modulates receptive field properties of relay cells in LGN.
Our own mechanistic models of the thalamocortical pathway are highly simplified linear rate-based models. While they provide valuable insight into receptive field structures, they are unable to explain the dynamics of signal processing in the early visual pathway, from retina via LGN to cortex. We are therefore developing a comprehensive model of the corticothalamic pathway (CoThaCo) aimed at the investigation of visual signal processing in the early visual pathway. The model is implemented using the Neural Simulation Tool NEST.
We are specifically interested in the interplay between pathways of different spatial and temporal resolution, and the role of feedback. Such a model requires a compromise between the desire for biological realism, and the constraints set by computer resources. Through the use of simpler neuronal models than those used by other researchers (Hill and Tononi, 2004), we are able to incorporate several thalamic pathways in our model network, while retaining a computational efficiency that allows the investigation of a range of stimulus paradigms. A sketch of the CoThaCo model is shown in the figure above.
When fully implemented and validated, the CoThaCo model will provide a unique substrate for in computo experiments on the early visual pathway. We will be at liberty to add, remove, or modify any parts of the system to test hypotheses about their role. We are at the same time able to record data from all neurons in our network simultaneously, granting us full control over the systems dynamics. It is our hope that this will further the understanding of the thalamocortical feedback loop significantly. The figure below shows the response of a partial implementation of the network to a "movie" of a walking robot.
Publications
Financial support
Our work on the development of CoThaCo and NEST would have been
impossible without financial support by