Why synchrony?

Figure 113.1 illustrates how relations among features can be encoded by conjunction-specific neurons within hierarchically structured feedforward architectures. Representing relations among components by conjunction units has two undisputed advantages. First, it permits rapid processing because it can be realized in feedforward architectures. Second, it is unambiguous because the response of a particular cell always signals the same relation (labeled line coding). However, if not complemented by additional, more dynamic and context-sensitive mechanisms for the definition of relations, this strategy poses problems. First, excessively large numbers of conjunction units are required to cope with the manifold intramodal and cross-modal relations between the features of real-world objects. Second, it is hard to see how the entirely new relations among the features of novel objects can be represented, as there cannot be an exhaustive repertoire of a priori specified conjunction units for all possible feature constellations. Third, unresolved problems arise with the specification of the nested relations that need to be defined to represent composite objects or scenes containing numerous objects. (For a more detailed review of the arguments, see Gray, 1999; Singer, 1999; von der Malsburg, 1999.)

Figure 113.1..  

Schematic wiring diagram of a hierarchically organized feedforward network that generates conjunction-specific neurons which respond selectively to different perceptual objects. Note that the neurons representing the faces and the vase receive input partially from the same feature-specific neurons.

These shortcomings can be overcome by assembly coding. Here individual cells signal only components of objects that may consist of rather complex conjunctions of elementary features—while the whole object is represented by the simultaneous responses of the respective component coding cells (Fig. 113.2). Individual neurons can then contribute at different times to the representation of different objects by forming ensembles with varying partners. This reduces dramatically the number of required conjunction units. It also solves the problem of representing novel relations, objects, and scenes because cells representing elementary features can be grouped dynamically in ever-changing constellations and then can represent as an assembly the particular combination of features characteristic of the novel object. However, this coding strategy requires a dynamic binding mechanism that can associate cells into assemblies and tag responses of distributed neurons as related once these have formed an assembly. Such dynamic binding mechanisms cannot be implemented in feedforward architectures, as they require additional reciprocal association connections and reentry loops.

Figure 113.2..  

Schematic wiring diagram of neuronal architectures serving the representation of perceptual objects by assemblies. Note that the assembly representing the vase shares neurons with the assemblies representing the faces. To ensure stability of the respective assemblies, additional reciprocal connections among neurons constituting an assembly are required (shaded regions) that bind responses of neurons belonging to the same assembly.

An unambiguous signature of relatedness is absolutely crucial in assembly coding because, unlike in labeled line codes, the meaning of responses changes with the context in which they are embedded. It needs to be ensured that the responses of the neurons that constitute an assembly are processed and evaluated together at subsequent processing stages and are not confounded with other, unrelated responses. In principle, this can be achieved by raising jointly and selectively the saliency of the responses belonging to an assembly, and there are three options: first, unrelated responses can be inhibited and excluded from further processing. Second, the discharge frequency of the selected responses can be enhanced. Third, the selected cells can be made to discharge in precise temporal synchrony. All three mechanisms enhance the relative impact of the selected responses and can therefore be used to tag them as related. Problems arise, however, when several assemblies need to be formed at the same time and need to recruit partly the same neurons—a condition that is likely to occur when a scene contains several objects that have subsets of features in common. This problem has been addressed as the so-called superposition problem and can only be resolved by segregating assemblies in time (Fig. 113.3). One option is to raise jointly the discharge frequency of cells belonging to the first assembly for an interval sufficiently long to permit readout by subsequent stages and subsequently to increase the discharge rate of the cells belonging to the second assembly, and so on. Another option is to label responses of cells belonging to an assembly by synchronizing the respective discharges. As synchronization raises the saliency of individual discharges, it defines relations with much higher temporal resolution than rate modulation and, in principle, permits multiplexing of assemblies, as suggested in Figure 113.3. Single-cell studies have provided robust evidence that attentional mechanisms use inhibition and selective enhancement of the discharge rate for response selection and grouping. However, there is little evidence so far from multielectrode studies that joint increases in discharge rate are used for response selection and grouping in the context of assembly coding. This may be due to the paucity of multielectrode studies searching for such phenomena. However, it is also conceivable that synchronization and temporal multiplexing are better suited for the definition of relations in assembly coding than joint rate increases.

Figure 113.3..  

Options for the solution of the superposition problem. Superposition problems arise if perceptual objects are present whose corresponding assemblies partly share the same neurons (upper box). In this case, different assemblies need to be segregated in time to avoid false conjunctions. One option is to raise successively the saliency of responses belonging to the respective assemblies by enhancing the discharge rate of the corresponding responses (lower box, option 1). An alternative solution is to enhance the saliency of responses belonging to a particular assembly by making the discharges of the respective neurons coincident in time (option 2). This permits rapid multiplexing of the different assemblies because coincidence can be evaluated within short time intervals, as it does not require temporal summation. Here it is assumed that the different assemblies alternate at intervals of approximately 25 msec. Note that this temporal structure can be, but does not have to be, obvious in the discharge sequences of individual neurons (channels 1 to 12) but that the spike density function of the population response shows an oscillatory modulation in the 40 Hz range. Note also that the constellation of neurons contributing spikes to the oscillatory population response changes from cycle to cycle.

In assembly coding processing, speed is limited essentially by the rate at which relational codes can be generated and read out. If relations are defined by joint rate increases, readout requires temporal summation of successively arriving excitatory postsynaptic potentials (EPSPs) and hence integration over a sequence of EPSPs. If integration intervals are shorter than the average interspike intervals, EPSPs do not summate effectively and rate increases cannot be detected. Synchronization, by contrast, exploits exclusively spatial and no additional temporal summation to raise the saliency of responses. Therefore, relations can be defined—at least in principle—with a temporal resolution corresponding to the duration of individual EPSPs. Thus, assemblies can be reconfigurated and multiplexed at much faster rate when synchronization is used as a tag of relatedness instead of joint rate enhancement. Another, potentially important advantage of using synchrony as a tag of relatedness is that relations can be specified independently of firing rate. Discharge rates depend on numerous variables, such as the physical energy of stimuli or the match between stimulus and receptive field properties. Response amplitudes are thus rather ambiguous indicators of relatedness. As synchrony can be modulated by temporal regrouping of discharges and thus can be varied independently of firing rates, synchronicity and rate can be used as orthogonal codes. Signals indicating the presence and the properties of visual features can thus be kept separate from signals indicating how these features are related.

Another advantage of using synchronization as a tag of relatedness is that synchronized input is transmitted with minimal latency jitter (Abeles, 1982; Diesmann et al., 1999). Thus, signatures of relatedness can be relayed with great reliability across processing stages, which contributes to reducing the risk of false conjunctions. Finally, synchronization enhances processing speed by accelerating synaptic transmission per se because synchronized EPSPs trigger action potentials with minimal delay.