Spike synchronization in feed-forward subnetworks of the cortex has been proposed to explain the precisely timed spike patterns observed in experiments. While the attractor dynamics of these networks is now well understood, the underlying single neuron mechanisms remained unexplained. Previous attempts have captured the effects of the highly fluctuating membrane potential by relating spike intensity f(U) to the instantaneous voltage U generated by the input. In this project we showed that f is high during the rise and low during the decay of U(t), demonstrating that the -dependence of f, not refractoriness, is essential for synchronization. Moreover, the bifurcation scenario is quantitatively described by a simple relationship. These findings suggest as the relevant model class for the investigation of neural synchronization phenomena in a noisy environment.

The balanced random network model attracts considerable interest because it explains the irregular spiking activity at low rates and large membrane potential fluctuations exhibited by cortical neurons in vivo. In this project we investigate to what extent this model is also compatible with the experimentally observed phenomenon of spike-timing-dependent plasticity (STDP). We observe that the experimental STDP data are well described by an update rule with a multiplicative dependence on the synaptic weight for depression and a power law dependence for potentiation. We show that this rule,when implemented in large, balanced networks of realistic connectivity and sparseness, is compatible with the asynchronous irregular activity regime. The resultant equilibrium weight distribution is unimodal with fluctuating individual weight trajectories and does not exhibit development of structure. We investigate the robustness of our results with respect to the relative strength of depression. We introduce synchronous stimulation to a group of neurons and demonstrate that the decoupling of this group from the rest of the network is so severe that it cannot effectively control the spiking of other neurons, even those with the highest convergence from this group.

Learning to predict future events and how to make choices is essential for the survival of every creature. One method specialized for prediction problems, which has been developed in the field of machine learning, is temporal-difference (TD) learning. However it is unclear how temporal-difference learning could be implemented in the brain. In this project we constructed a spiking neural network model that implements actor-critic temporal-difference learning. The fundamental units of the actor-critic model, the policy and the value function, are represented by synaptic weights. We derive a quantitative correspondence of the synaptic weight parameters to the corresponding variables in the discrete time computer algorithm. The synaptic learning mechanisms combine local plasticity rules motivated by biological findings with a global reward signal. We tested the learning behavior of the neural network in navigating to a reward in a two-dimensional grid-world environment. The neural network is able to accurately evaluate the quality of a state and adapts its behavior. Performance speed and stability of the neural network are similar to the discrete time counterpart. This work shows that a network of spiking neurons realizing TD learning is capable of solving complex tasks.

The non-uniform and specific connectivity of the local cortical network provides the structural basis for the function of the constitutive information-processing unit usually referred to as cortical column, cortical module or canonical microcircuit. Large-scale simulations with layer-specific connectivity allow us to investigate the dynamical implications of substantial deviations from uniform connectivity. We use published data to extract and integrate the information required to construct a multi-layered neocortical network model. On the basis of this information we explore the dynamical properties induced by the layer-specific connectivity using simulations of a local cortical module consisting of 100,000 neurons. In so doing we are able to check for the consistency of the known structure of the neocortex and its observed network activity, linking structure to dynamics. This work is partially supported by EU Grant 15879 (FACETS).

Spike-timing dependent plasticity (STDP) has received considerable attention in recent years due to its potential relevance for shaping the functional connectivity structure of neural networks dependent on the correlated activity of the pre- and postsynaptic cells. An implementation of STDP in VLSI spiking neural network models would enable researchers to utilize the computational power of the hardware for tasks which are difficult to solve with software approaches (VLSI: very large scale integration).
In the context of the FACETS project [1], we investigate the effects of a limited resolution for storing synaptic weights on their dynamics and on the dynamics of neural activity. For this purpose we use a NEST-based partial software emulation of the STDP VLSI-implementation that is currently developed in the FACETS project [2]. Our analysis provides insights into the capabilities of current and future VLSI implementations of STDP. This collaboration with the electronic vision(s) group at the Kirchhoff Institute for Physics in Heidelberg, Germany [3] supports the development of hardware that is able to reproduce synaptic weight trajectories as evoked by biologically relevant spike statistics.

Correlations between individual channels have been observed at signal levels ranging from fMRI Bold over the local field potential down to the spike count. In some approximation these signals can be considered as superpositions of spike trains filtered by biophysical components of the brain and the measurement process. However, it is largely unknown how the correlation structure is altered by this filtering and what the consequences for the dynamics of the system and for the interpretation of measured correlations are. In this study we focus on linearly filtered spike trains and particularly consider correlations caused by overlapping presynaptic neuron populations. We demonstrate that correlation functions and statistical second-order measures like the variance, the covariance and the correlation coefficient generally exhibit a complex dependence on the filter properties and the statistics of the individual presynaptic spike trains. Both contributions can significantly modulate modulating the apparent interaction strength between neurons or neuron populations. In contrast, the coherence measure in many applications allows a filter independent quantification of correlated activity. In different network models we discuss the estimation of network connectivity from the high-frequency coherence of simultaneous intracellular recordings of pairs of neurons.

The modeling demands in computational neuroscience require software tools able to cope with modern high-performance clusters and high-end supercomputers. To this end we develop in collaboration with several other institutes and the Honda Research Institute Europe the Neural Simulation Tool NEST. NEST allows distributed simulation of large-scale neural networks showing supralinear scaling with the number of computing cores. The upcoming release NEST2 efficiently distributes simulations of up to 10⁶ neurons. In order to keep up with the current development in processor technology NEST2 is optimized for next-generation multicore architectures. It employs a hybrid message-passing and multi-threading technology implicitly, i.e. the user needs not to be aware of the underlying complexity. Furthermore NEST2 allows exact integration with continuous spike times. Most recent developments make the user-friendly python-based interface PyNEST available.

This work is carried out in collaboration with the NEST Initiative.

Explicit network realizations of system-level theories of brain function are needed to address their biological plausibility, and to allow relevant experiments to be designed. Here, we consider the concept of compositionality, i.e. the representation of complex objects by the combination of representations of its constituting parts. A hierarchic structure of synfire chains has been proposed as the neural substrate for compositionality. We demonstrate that such a structure is indeed capable of modelling compositionality not only of perception but also of the production of behavior. As an experimentally relevant example we address the generation of arm movements. A small number of synfire chains is sufficient to generate arbitrary planar point-to-point trajectories on the basis of simple motion primitives. Moreover, when the network is constructed in a closed-loop fashion, autonomous scribbling-like trajectories are generated. Thus, in this model we can study the relationship between behavior and neuronal activity. The results can then be compared to the relationships found in experiments.

In particular the project investigates:

- the neural mechanisms for binding and compositionality

- a prototypical but complete composition machine

- the generation of scribbling trajectories on the basis of movement primitives

This work is a component of project DIP F1.2 Compositionality: Neuronal basis of complex behaviour in collaboration with Moshe Abeles, Tamar FLash, Mina Teicher, Theo Geisel, and Michael Herrmann.