next up previous
Next: Robust Estimation Up: Construction by Coherence-Detection Previous: Marking by a synchronous

Preknowledge coded in synaptic links

Within the context of dynamical coherence detection, the synaptic link structure has a different computational purpose. In any given cortical area, specific neural units will be sometimes co-activated by incoming stimuli, while others will never be. Which units will be co-activated and which not will depend on the statistical nature of the input signals. A synaptic link structure can adapt to this statistical structure of the incoming neural signals by increasing the connection strength between units which are often co-active. With such an adapting link structure, new, but similar stimuli will be able to recruit more neural units into a coherence cluster, which can stabilize the estimates derived by the coherence detection process.

For example, in the auditory cortex co-activation of neurons responding to higher harmonics of a single frequency is to be expected quite often, since most auditory stimuli in the real world are composed of a base frequency plus higher harmonics. Neurons responding to non-harmonics might also be sometimes co-active - but only by chance, and not in general. In the same way, disparity sensitive neurons in the visual cortex will sometimes be co-activate if they cover the same view-direction with their receptive fields. For neurons responding to different view-directions, co-activation will be purely coincidental.

In this way, an adaptive synaptic link structure can adopt to the main characteristics of the input signals. The important point is that under the weak-coupling paradigm, a given synaptic link is not enough to group neurons together. The synaptic link structure of a network presents only possibilities of interaction, but at any given moment most of these links will not be dynamically utilized. Usually, only one of all possible groupings supported by the synaptic link structure will be instantiated by an actual stimulus.

Fig. 2 shows an example of such a stimulus-dependent utilization of synaptic links. In this simulation, all neurons of the network have synaptic links with each other. However, for a given stimulus, only a small subgroup of all the neurons synchronize; which one depends on the current stimulus value.

Conceptually, dynamical coherence detection and synaptic link modifications operate on two different time scales. On a short time scale, coherence detection analyses the current stimulus situation; on a much longer time scale, synaptic link modifications record the long-time statistics of the incoming signals.


next up previous
Next: Robust Estimation Up: Construction by Coherence-Detection Previous: Marking by a synchronous

2000-11-20