As soon as the neurons within the coherence cluster fire synchronously, the characteristics of the total output current of the network changes (Fig. 3). With weak coupling switched off, the fluctuating amplitude of the output current only reflects random coincidences of spikes (time in Fig. 3). Turning the weak coupling on between neurons (at times ) the synchronization of the coherence cluster leads to a strong periodic modulation of the output current, easily detectable by subsequent layers. Since the depth of this modulation depends on the relative percentage of neurons in the coherent cluster, it is a good indication of the validity of the estimate.
The marking of neural groups by a synchronous time code is an old idea [18,19] and has been subject to various modeling attempts over time. But the dynamical coherence detection scheme proposed here differs in important aspects from previous ideas.
Most of the other schemes use a much stronger coupling between neurons, compared to the weak-coupling paradigm developed here. But with stronger coupling, the whole network of oscillators quickly synchronizes, and all information about the stimulus value it codes is lost.
Avoidance of full synchronization defines an upper limit for the term ``weak-coupling'' introduced here. Dynamical coherence detection is only possible within a range of intermediate coupling strengths, located between the two extremes of ``independent oscillators'' and ``fully synchronized network''.
Too strong coupling between oscillatory units can further result in complicated network behaviour, including chaotic dynamics and oscillator death [7,20]. With strong coupling the internal characteristics of the oscillatory units become important - and they can no longer be described as interacting phaseoscillators.
Many of the network models proposed in the literature use oscillatory elements with the same frequency of oscillation, facilitating locking of all oscillators. Stimulus values are coded in all these models not by intrinsic oscillator properties, but by the place of the activated elements in the networks. Most of the time, a binary valued stimulus is used as input, to switch the oscillatory elements ``on'' or ``off''. Connected by preset [21,22,23,24,25,26,27,28] or prelearned [29,30] synaptic couplings, the activated and connected oscillators quickly synchronize; the non-activated units of the network remain in the quiescent state.
Thus, synchronization is used in these approaches not as a computational tool, but only to mark a preselected group of oscillatory elements. Other efforts to model feature-binding [19,31,32,33,34,35,36] rely on fast synaptic links, modified by activity-driven learning rules. In effect, these learning rules couple only neural units with a common high activity together. Thus, from a computational point of view, models utilizing fast synaptic links are similar to models utilizing oscillatory units being switched ``on'' or ``off''. Primary computational units are the dynamical synapses, and synchronization is used only to mark the group of synaptically coupled neurons.
Dynamical coherence detection uses quite different concepts. Here, stimulus values are continously represented as an internal property of the oscillatory units, namely the frequency of oscillations, and synchronization is used as computational tool to detect the coherence cluster. Synaptic links are necessarily weak, barely changing the normal, stimulus-driven activity of single units, and fixed, at least on short timescales.