Jan Storck

Temporal clustering with spiking neurons and dynamic synapses: towards technological applications

We apply spiking neurons with dynamic synapses to detect temporal patterns in a multi-dimensional signal. We use a network of integrate-and-fire neurons, fully connected via dynamic synapses, each of which is given by a biologically plausible dynamical model based on the exact pre- and post-synaptic spike timing. Dependent on their adaptable configuration (learning) the synapses automatically implement specific delays. Hence, each output neuron with its set of incoming synapses works as a detector for a specific temporal pattern. The whole network functions as a temporal clustering mechanism with one output per input cluster. The classification capability is demonstrated by illustrative examples including patterns from Poisson processes and the analysis of speech data.

Speech recognition with spiking neurons and dynamic synapses: a model motivated by the human auditory pathway

Abstract We present a biologically motivated system for auditory signal processing and speech perception. In a novel approach, a model of the human cochlea is combined with a network of spiking neurons and dynamic synapses. The cochlea model continuously transforms the raw acoustic waveforms to multi-channel spike patterns. The frequency and temporal information is thereby coded similar to the auditory nerve. A network of spiking neurons and dynamic synapses is able to learn these spatio-temporal patterns by establishing characteristic delay structures in its connections due to synaptic plasticity.

Nonlinear independent component analysis and multivariate time series analysis

Abstract We derive an information-theory-based unsupervised learning paradigm for nonlinear independent component analysis (NICA) with neural networks. We demonstrate that under the constraint of bounded and invertible output transfer functions the two main goals of unsupervised learning, redundancy reduction and maximization of the transmitted information between input and output (Infomax-principle), are equivalent. No assumptions are made concerning the kind of input and output distributions, i.e. the kind of nonlinearity of correlations. An adapted version of the general NICA network is used for the modeling of multivariate time series by unsupervised learning. Given time series of various observables of a dynamical system, our net learns their evolution in time by extracting statistical dependencies between past and present elements of the time series. Multivariate modeling is obtained by making present value of each time series statistically independent not only from their own past but also from the past of the other series. Therefore, in contrast to univariate methods, the information lying in the couplings between the observables is also used and a detection of higher-order cross correlations is possible. We apply our method to time series of the two-dimensional Henon map and to experimental time series obtained from the measurements of axial velocities in different locations in weakly turbulent Taylor-Couette flow.

Learning spatio-temporal stimuli with networks of spiking neurons and dynamic synapses

Abstract A network of spiking neurons and dynamic synapses is introduced to yield a mechanism for learning spatio-temporal stimulus patterns. Integrate-and-fire postsynaptic neurons receive input spike trains from multiple dynamic synapses. The synaptic dynamics is based on exact pre- and postsynaptic spike timing and exhibits short-term facilitation and depression. In addition, dependent on their adaptable long-term configuration (learning) the synapses automatically implement specific delays in their peak response. Each postsynaptic neuron with its set of incoming synapses gets tuned to a specific spatio-temporal pattern. The whole network is capable of discriminating between stimuli with one output per learned stimulus type.

Spike-Based Hebbian Learning for Stimulus Discrimination

We analyse the impact of Hebbian Learning on a network of spiking neurons. The network consists of pyramidal cells each of which is connected to other pyramidal cells and also locally to an inhibitory stellate cell. The neurons are described by the spike response model (SRM). The network can be driven by different classes of sensorial stimulus. The learning results in stimulus dependent spatio-temporal spike patterns, which are given by clusters of synchronously firing neurons, such that the stimuli can be easily discriminated. A two-picture experiment serves as illustration.