Oscillations, Synchrony, and Neuronal Codes

views updated


Cognitive systems have to explore a huge combinatorial space when searching for the consistent relations among features that define a perceptual object. Thus, mechanisms are required that permit rapid analysis and representation of relations between the responses of neurons whose activity signals the presence of particular features. A common and well-documented strategy for the binding of distributed responses is the implementation of conjunction-specific neurons that receive convergent input from elementary feature detectors and therefore respond selectively to the conjunctions of the respective features. This process is known as labeled line coding. However, this coding strategy, if not complemented by an additional binding mechanism, meets with a number of problems. First, large numbers of conjunction units are required for the exhaustive representation of the manifold intra-and cross-modal feature constellations of real-world objects. Second, it is hard to see how novel objects and hence entirely new relations among features can be recognized and represented because this would require rapid reconfiguration of input connections to previously uncommitted cells. Third, unresolved problems arise with the representation of the nested relations among the components of composite objects such as visual scenes or sentences (Singer, 1999; von der Malsburg, 1999).

A complementary strategy is needed, therefore, that permits a more flexible definition of relations than can be achieved with hard-wired conjunction units. As proposed by Donald Hebb (1949) and later elaborated by others, that complementary strategy is assembly coding. The assumption is that only components of objects—which may consist of rather complex conjunctions of elementary features but may be common to different objects—are represented by individual cells. The presence of the whole object seems to be signaled by the simultaneous responses of the ensemble of cells responding to the components. Thus, in ensemble coding individual neurons can contribute at different times to the representation of different objects by forming ensembles with varying partners. A neuron tuned to a particular component can then contribute to the representation of all objects containing this particular component and neurons representing elementary features can be recombined in ever-changing constellations to represent novel objects.

How to Tag Responses as Related?

Numerous theoretical studies have addressed the question how assemblies can organize themselves through cooperative interaction within associative neuronal networks. Here we focus on the problem of how cells can be tagged as related when they are grouped into an assembly. An unambiguous signature of relatedness is absolutely crucial for assembly codes because, in contrast to labeled line codes, the meaning of responses changes with the interpretive context, thus rendering false conjunctions deleterious for object recognition. The required mechanism must assure 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, such a process can be achieved by raising jointly and selectively the saliency of the responses belonging to an assembly.

There are three ways to achieve this goal: the inhibition and exclusion of responses from further processing, the enhancement of the discharge frequency of the selected responses, and the inducement of precise synchrony of the selected cells. All three mechanisms enhance the relative impact of the selected responses and can therefore be used to tag them as related. Single-cell studies have provided robust evidence that the first two mechanisms play a crucial role in the selection and grouping of responses.

The simultaneous organization of assemblies sharing common subsets of neurons is precluded by the impossibility of knowing which of the shared neurons would belong to which assembly. Hence, the frequent overlapping of assemblies necessitates a temporal segregation through multiplexing. Processing speed is then limited essentially by the rate at which different assemblies can follow one another and thus by the temporal resolution of the mechanism that labels responses as related. Through synchronization—the temporal regrouping of spikes—the saliency of responses can be modulated with higher temporal resolution than with tonic changes in the firing rate. Synchronization exploits exclusively spatial and no additional temporal summation, and therefore this tagging mechanism can operate in principle with a temporal resolution at the level of individual spikes. Using synchronization as a complementary mechanism for the definition of relations also permits the possible advantage of specification of relations independently of the firing rate. The discharge rates of neurons depends on numerous variables such as the physical energy of stimuli or the match between stimulus and receptive field properties, and it may not always be obvious how these modulations of response amplitude can be distinguished from those signaling the relatedness of responses. Not all strong responses are necessarily related.

Synchrony as a Code for the Definition of Relations

Some researchers have suggested that the cerebral cortex imposes a temporal micro-structure on otherwise sustained responses and uses this temporal patterning to express through synchronization the degree of relatedness of the responses (Singer, 1999; Engel, Fries, and Singer, 2001). This suggestion ensued from the following discoveries: cortical neurons often engage in synchronous oscillatory activity that is not stimulus-locked but caused by internal interactions (Gray and Singer, 1989); neurons distributed both within and across cortical areas can synchronize their discharges within a millisecond; and synchronization probability reflects common Gestalt-criteria of perceptual grouping.

If internally generated synchronization were to serve as a signature of relatedness, it would need to meet several criteria: First, its precision should be in the millisecond range to match the temporal windows for effective spatial summation and Hebbian modifications. Second, it must be possible to generate and dissolve episodes of synchronous firing at a rate fast enough to be compatible with known processing speed. Third, synchronized activity must be more effective than nonsynchronized activity in driving cells in target structures because it can only serve as a tag of relatedness if it enhances the saliency of the synchronized responses. Fourth, there should be correlations between the occurrence of synchronization patterns and perceptual or motor processes. These predictions are supported by experimental evidence (Gray, 1999; Singer, 1999). Here we shall review only a selection of studies addressing the last postulate.

Attention and Response Selection

Spike synchronization and the often concomitant oscillatory patterning of responses in the β- and γ-frequency range are particularly well expressed when the brain is in an activated state—when the EEG is desynchronized and exhibits high power in the β- and γ-frequency range. Such EEG patterns are characteristic of the aroused, attentive brain, suggesting a role for synchronization in cognitive processes. This suggestion is supported by numerous observations in both animals and human subjects that synchronous oscillations in the γ-frequency range and their synchronization become more prominent during states of focused attention or when subjects are engaged in cognitive tasks that put strong demands on feature binding or short-term memory functions (Tallon-Baudry and Bertrand, 1999; Engel, Fries, and Singer, 2001). Multielectrode recordings from awake cats and monkeys trained to perform discrimination tasks indicate that attentional mechanisms enhance neuronal synchrony in anticipation of the expected task. There is an increase in oscillatory activity in the gamma-frequency range that is associated with increased coherence between the spontaneous discharges of cells and the oscillations of the local field potential (Roelfsema, Engel, König, and Singer, 1997). These attentional effects appear to be selective and confined to cortical areas and sites that need to be engaged for the execution of the anticipated tasks (Fries, Reynolds, Rorie, and Desimone, 2001).

A close correlation between response synchronization and stimulus selection has been found in experiments on binocular rivalry that were performed in strabismic animals (Fries, Schröder, Singer, and Engel, 2002). Because of experience-dependent modifications of processing circuitry, perception in strabismic subjects always alternates between the two eyes. We have exploited this phenomenon of rivalry to investigate how neuronal responses that are selected and perceived differ from those that are suppressed and excluded from supporting perception (see Figure 1). A close and highly significant correlation was observed between changes in the strength of response synchronization and the outcome of rivalry. Cells mediating responses of the eye that won in interocular competition increased the synchronicity of their responses upon presentation of the rivalrous stimulus to the other, losing eye, whereas the reverse was true for cells driven by the eye that became suppressed. Surprisingly, there were no consistent modifications of the amplitude of responses. It is only at later processing stages that the poorly synchronized responses to the suppressed stimuli fail to elicit suprathreshold responses and that cells respond only to the selected stimulus (Leopold and Logothetis, 1996).

Synchronization and Feature Binding

The hypothesis that internal synchronization of discharges groups responses for joint processing predicts that synchronization probability should reflect some of the basic Gestalt criteria according to which the visual system groups related features during scene segmentation. A series of studies provided evidence that neurons distributed across different columns within the same or different visual areas and even across hemispheres synchronize their responses with almost no phase lag when activated with a single contour but fire independently when stimulated simultaneously with two different contours. This pattern suggests that synchronization results from a context-dependent selection and grouping process. The probability and strength of response synchronization reflect indeed some of the elementary Gestalt-criteria that underly perceptual grouping. Stimulus configurations that comply with criteria such as continuity, proximity, and similarity in the orientation domain, colinearity, and common fate evoke synchronized responses with higher probability than configurations that are devoid of groupable features (Singer, 1999).

Researchers have observed an especially close correlation between neuronal synchrony and perceptual grouping in experiments with plaid stimuli. These stimuli are well suited for the study of dynamic binding mechanisms because minor changes of the stimulus cause a binary switch in perceptual grouping. Two superimposed gratings moving in different directions (plaid stimuli) may be perceived either as two surfaces, one being transparent and sliding on top of the other (component motion), or as a single surface, consisting of crossed bars that moves in a direction intermediate to the component vectors (pattern motion). If this grouping of responses is initiated by selective synchronization, three predictions must hold (see Figure 2): First, neurons that prefer the direction of motion of one of the two gratings and have colinearly aligned receptive fields should always synchronize their responses because they respond always to contours that belong to the same surface. Second, neurons that are tuned to the respective motion directions of the two gratings should synchronize their responses in case of pattern motion because they then respond to contours of the same surface, but they should not synchronize in case of component motion because their responses are then evoked by contours belonging to different surfaces. Third, neurons preferring the direction of pattern motion should also synchronize only in the pattern and not in the component motion condition.

Cross-correlation analysis of responses from cell pairs distributed either within or across areas 18 and the posterior medio-lateral suprasylvian sulcus (PMLS) of the cat visual cortex confirmed all three predictions. Cells synchronized their activity if they responded to contours that are perceived as belonging to the same surface (Castelo-Branco, Goebel, Neuenschwander, and Singer, 2000) (see Figure 2C). Dynamic changes in synchronization could, thus, serve to encode in a context-dependent way the relations among the simultaneous responses to spatially superimposed contours and thereby bias their association with distinct surfaces.


There is evidence that neuronal networks can synchronize neuronal discharges with a precision in the millisecond range and appear to exploit this ability for at least three purposes: first, for the precise signaling of temporal features across processing stages; second, for the selection of responses; and third, for the definition of relations among distributed responses with high temporal resolution. This selection and binding mechanism is best for ensemble coding because it meets the requirement for flexible and rapid binding of distributed responses in ever-changing constellations. Assembly coding, in turn, appears necessary in order to cope with the representation of the astronomical number of possible relations among features describing real-world objects.

It appears, then, as though the cerebral cortex applies two complementary coding strategies: an explicit representation of features and their conjunctions in the tuned responses of individual specialized neurons or populations of such neurons and an implicit representation of conjunctions of such explicitly coded contents in dynamically associated assemblies. The first strategy seems to apply to the representation of a limited set of features and some of their conjunctions and is in all likelihood reserved for items that occur very frequently and/or are of particular behavioral importance. The second strategy seems to be reserved for the representation of novel objects and of all those items for which an explicit representation cannot be realized, either because the explicit representation would require too many neurons or because the contents to be represented are too infrequent to warrant the implementation of specialized neurons.



Castelo-Branco, M., Goebel, R., Neuenschwander, S., and Singer, W. (2000). Neural synchrony correlates with surface segregation rules. Nature 405, 685-689.

Engel, A. K., Fries, P., and Singer, W. (2001). Dynamic predictions: Oscillations and synchrony in top-down processing. Nature Reviews Neuroscience 2, 704-716.

Fries, P., Reynolds, J. H., Rorie, A., and Desimone, R. (2001). Modulation of oscillatory neuronal synchronization by selective visual attention. Science 291, 1,560-1,563.

Fries, P., Roelfsema, P. R., Engel, A. K., König, P., and Singer, W. (1997). Synchronization of oscillatory response in visual cortex correlates with perception in interocular rivalry. Proceedings of the National Academy of Sciences of the United States of America 94, 12,699-12,704.

Fries, P., Schröder, J.-H., Singer, W., and Engel, A. K. (2002). Oscillatory neuronal synchronization in primary visual cortex as a correlate of perceptual stimulus selection. Journal of Neuroscience 22, 3,739-3,754.

Gray, C. M. (1999). The temporal correlation hypothesis of visual feature integration: Still alive and well. Neuron 24, 31-47.

Gray, C. M., and Singer, W. (1989). Stimulus-specific neuronal oscillations in orientation columns of cat visual cortex. Proceedings of the National Academy of Sciences of the United States of America 86, 1,698-1,702.

Hebb, D. O. (1949). The Organization of Behavior. New York: John Wiley.

Leopold, D. A., and Logothetis, N. K. (1996). Activity changes in early visual cortex reflect monkeys' percepts during binocular rivalry. Nature 379, 549-553.

Roelfsema, P. R., Engel, A. K., König, P., and Singer, W. (1997). Visuomotor integration is associated with zero time-lag synchronization among cortical areas. Nature 385, 157-161.

Singer, W. (1999). Neuronal synchrony: A versatile code for the definition of relations? Neuron 24, 49-65.

Tallon-Baudry, C., and Bertrand, O. (1999). Oscillatory gamma activity in humans and its role in object representation. Trends in Cognitive Sciences 3, 151-162.

von der Malsburg, C. (1999). The what and why of binding: the modeler's perspective. Neuron 24, 95-104.