Projects
From NEST
Contents |
Cortex models
Effects of realistic network size and connectivity
Research indicates that higher brain functions emerge from large and complex cortical networks and their interactions. Assuming a biologically realistic level of connectivity, each cortical neuron should have of the order of 10,000 presynaptic inputs. In order to ensure that the network is sufficiently sparse, a connection probability of 0.1 should be assumed, resulting in a minimal network size of 105. Such a network contains of the order of 109 synapses, and as such has proved to be beyond the memory capacity of the computers available to many researchers. Simulation studies mimicking a reasonable connectivity were so far restricted to network sizes of up to 10,000 neurons, which is the lower boundary for a minimal, functionally relevant cortex volume. On the other hand, theoretical studies usually deal with infinitely large systems. It turned out, that both levels of description can lead to qualitatively different results. Scaling of neural networks (decreasing the system size while increasing the connection strengths) is not possible. Here, we try to overcome the gap between theory and numerics with the help of NEST which allows to build cortex-like networks with a realistic connectivity for up to 1,000,000 neurons.
(Diesmann 1999, Diesmann 2001, Gewaltig 2001, Tetzlaff 2002, Aviel 2003, Mehring 2003, Tetzlaff 2003, Tetzlaff 2004, Morrison 2004, Tetzlaff 2005)
Network topology
Random network models have been a popular tool for investigating cortical network dynamics. On the scale of roughly one cubic millimeter of cortex, containing about 100,000 neurons, cortical anatomy suggests a more realistic architecture with a spatially structured connectivity. Here, we address the question, how network dynamics is affected by the properties of the connectivity matrix. Candidate architectures include randomly connected networks, regular networks with a fixed number of neighbours per node, as well as "small world" and "scale-free" networks, which lie somewhere in between those two extremes. Last but not least, we will examine networks with structural parameters obtained by neuroanatomical methods, to find out how they deviate from the more abstractly defined substrates, and what the functional implications of these deviations are.
(Mehring 2003, Steimer 2004)
Learning in Plastic Cortical Systems
A useful model for small volumes of the cortex is that of a network where excitation is balanced by inhibition, producing irregular firing patterns at low rates. In the case of static synapses, such networks have been thoroughly studied and are well understood. Here, we investigate to what extent the STDP update rules are compatible with a balanced network model. If they are compatible, it should be possible to construct a balanced recurrent network with appropriate activity statistics, where the distribution of synaptic weights is centred around low values, so that the network is sensitive to correlations in its input. Such systems are less tractable than those with static synapses for three reasons:
- In the general case rates and weight distribution are interdependent.
- In a recurrent network, higher order correlations occur due to the finite network size. The extent and effect of these correlations are still hard to quantify.
- The timescale over which a distribution of weights relaxes to its stationary state is very long (100-1000s), so network simulations take a long time to perform.
From this project we expect to establish whether the concepts of balanced networks and synaptic plasticity are consistent, and obtain insight into the complex interaction between system level and cellular level dynamics.
(Morrison 2004b)
Rate-Correlation dynamics
The analysis of the spatial and temporal structure of spike cross-correlation in experimental data is an important tool in the exploration of cortical processing. Recent theoretical studies investigated the impact of correlation between presynaptic inputs on the spike rate of a postsynatpic neuron and the effect of input correlation on the output correlation of pairs ofneurons. Here, this knowledge is combined to a model simultaneously describing the dynamics of rate and correlation, allowing for an interpretation of its constituents in terms of network activity.
(Aviel 2003, Tetzlaff 2003, Schöner 2005)
Synfire chains
Cortical activity in vivo is characterised by asynchronous irregular firing of the neurons at low rates. In addition, the cortex exhibits precise spatio-temporal spike patterns in relation to the experimental protocol. Since individual synaptic events are usually weak, reliable and precise spike initiation requires a well-timed cooperation between several neurons. The concept of the "synfire chain" was introduced as a toy model to study how such a cooperative spiking can lead to the observed precise spike patterns even in the presence of a considerable amount of noise. Several approaches have been used to show that synchronous volleys of spikes can indeed stably propagate through the network for a wide range of parameters. However, in further studies we show that the embedding of such feed-forward sub-systems into a "cortical" surrounding can easily destabilise the system. Stable embedded synfire activity is therefore still a challenge.
(Diesmann 1999, Gewaltig 2001, Tetzlaff 2002, Aviel 2003, Mehring 2003, Tetzlaff 2003, Tetzlaff 2004)
Simulation technology
Distributed computing
The availability of efficient and reliable simulation tools is one of the mission critical technologies in the fast moving field of computational neuroscience. The large number of elements (neurons) combined with the high connectivity (synapses) of the cortical network and the specific type of interactions impose severe constraints on the explorable system size which have previously been hard to overcome. We present a collection of new techniques combined to a coherent simulation tool removing the fundamental obstacle in the computational study of biological neural networks: the enormous number of synaptic contacts per neuron. Distributing an individual simulation over multiple computers enables the investigation of networks orders of magnitude larger than previously possible. The software scales excellently on a wide range of tested hardware, so it can be used in an interactive/iterative fashion for the development of ideas, and results can be produced quickly even for very large networks. In contrast to earlier approaches, a wide class of neuron models and synaptic dynamics can be represented.
Partially funded by the Volkswagen Foundation, GIF, BIF, BMBF-DIP F1.2, and DAAD/NFR 313-PPP-N4-lk.
(Morrison 2004)
Off-grid events in discrete time simulation
The high convergence in cortical systems (10,000 input synapses) excludes an event driven simulation scheme where spike times can be handled as continuous variables. Thus, simulations with NEST are performed in discrete time. In a naive implementation, any elicited spike is forced onto the simulation time grid. Artificial synchronisation is one of the most serious consequences. In particular, for studies concerned with the correlation structure of the network activity, the ability to represent precise spike times is desirable. For this purpose, we 're developing new algorithms intended to unify discrete time simulation schemes with those basing on the interaction by continuous time events.
Partially funded by DAAD/NFR 313-PPP-N4-lk, BMBF-DIP F1.2, GIF, and the Honda Research Institute.
(Hake 2003, Straube 2004, Morrison 2005)
