The Bernstein Center Freiburg

Informal Seminar
Amelia Waddington

School of Computing
University of Leeds, UK

Local Spike-Time-Dependent Plasticity Rules Generate Feed Forward Structure in Neural Networks
Thursday, October 21, 2010

14:00 h sharp
Library, 1st floor
Bernstein Center
Hansastr. 9a
The ability of the brain to process information quickly, reliably and reproducibly is an important challenge for computational neuroscience. One architecture that has been proposed for its generation of precisely timed spike patterns is synfire chains: multi-layered, feed forward structures, characterised by effectively synchronous firing in each layer. Recordings consistent with such structures have been reported in a number of neural systems including song bird HVC, motor cortex of monkeys, visual cortex of cats and auditory cortex of rats. Understanding possible mechanisms that could support the development of synfire chains from initially homogeneously connected neural assemblies has been the subject of increasing recent attention in the computational neuroscience community.
Theoretical studies have shown that it is possible to develop synfire chains in populations of excitatory neurons by combining synapse specific plasticity rules (such as spike time-dependent-plasticity) with more global neuronal plasticity rules (e.g., pruning or hetero-synaptic plasticity rules that effectively cap the number of possible pre-synaptic and/or post-synaptic connections for an individual neuron).
Here, we ask whether synfire chains can develop in the absence of neuronal capping rules. We use a spike time-dependent-plasticity rule based on recent experimental observations. We show that to achieve robust, stable and scalable synfire chains, additional neuronal adaptation rules are indeed needed, but that depending on the form of the synaptic plasticity, these need not be capping rules. Removing the need for strong neural capping rules significantly broadens the range of network structure and allows for the modulation of the the number of layers and their relative sizes. We characterise the emergent network structures as a function of stimulus properties and the synaptic learning rate. Results are presented both for a computational simulation model and for a reduced analytical model describing the initial layer formation as a bounded random walk.
The talk is open to the public. Guests are cordially invited!