Cerveau et cognition: temporal Aspects of Information Processing: From Cellular Mechanisms to Perception.
Symposium organized by Yves Burnod and Simon Thorpe, 23 to 29 August 2000
Jean BULLIER (Faculté de médecine de Rangueil, Toulouse), Yves BURNOD (Institut de neuroscience de Vision), Jacques DEMONGEOT (Université J.-Fournier, La Tronche), Alain DESTEXНE (Unité de neurosciences intégratives et computationnelles, C.N.R.S., Gif-sur-Yvette) Reinhard ECKHORN (Université de Marburg, Allemagne ),Yves FREGNAC (Unité de neurosciences intégratives et computationnelles, C.N.R.S., Gif-sur-Yvette) Christian KEYSERS (University of St-Andrews, Écosse), Remy LESTIENNE (Institut des neurosciences, Université Paris-6), David McCORMICK (Neurobiology Department Yale University School of Medicine, New Haven, USA), Klaus PAWELZIK (Université de Brême, Allemagne ), Barry RICHMOND (Laboratoire de neuropsychologie N.I.M.Н., Bethesda, USA), David SHEINBERG (Department of Neurosciences, Brown University, Providence, USA), Catherine TALLON-BAUDRY (I.N.S.E.R.M., Lyon), Alex TНOMSON (University College, Londres), Simon TНORPЕ (Faculté de médecine de Rangueil, Toulouse), Roger TRAUB (University of Birmingham, Grande-Bretagne), Rufin Van RULLEN Faculté de médecine de Rangueil, Toulouse).
The original idea for organising a meeting on temporal aspects of information processing in the brain, from cellular mechanisms to perception came from Agnes Babloyantz. Agnes had long been interested in theoretical aspects of brain function, and she was one of the first to аpply techniques from non-linear dynamics to the analysis of brain activity. The analysis of the temporal dynamics of information processing has now been performed at the different levels of organisation of the brain, molecules, neurones, neural networks, neural structures, perception and cognition. This work has been done in parallel by different scientists focused on different levels of brain organisation. Agnes wanted to promote interactions between these scientists, to better understand how processes occurring at the cellular level can be responsible to higher level functions, such as perception. Agnes was very enthusiastic to organise a meeting in « Les Treilles », and we were very happy to participate in such a rare occasion.
Sadly, Agnes passed away in March 2000. On behalf of all the participants, we would like to express our sincere gratitude. We dedicated the meeting to her mеmory.
During the meeting, there was a good mix of experimental work and theory which attempt to answer four basic questions:
- What are the cellular mechanisms responsible for the neuronal temporal dynamics in visual information processing?
- What is the functional role of the temporal variability observed in the neuronal signals during visual information processing?
- What is the functional role of oscillations and synchronies observed in the neuronal signals during visual information processing?
- What is the neural basis of rapid image processing of natural scenes?
1- What are the cellular mechanisms responsible for the neuronal temporal dynamics and information processing?
Detailed neurophysiological studies at the cellular level, have revealed a set of complementary cellular mechanisms that are responsible for the temporal aspects of neuronal activity and information processing in the brain.
David McCormick from the Yale University School of Medicine provided some fascinating insights into the « Cellular mechanisms of cortical dynamics ». In particular, he concentrated on the cellular mechanisms underlying contrast adaptation, a psychophysical phenomenon whose neuronal basis appears to reside largely in the primary visual cortex. McCormick’s team has investigated the underlying mechanisms using intracellular recording and current injections. His conclusion is that the adaptation of neuronal responsiveness to varying levels of contrast is mediated largely through the control of the membrane potential by intrinsic neuronal mechanisms. In particular, ionic potassium channels are highly sensitive to the previous temporal patterns of membrane potential, and in turn, modify the sensitivity of the neurones to the next synaptiс inputs by changing the membrane conductance. This feedback mechanism, internal to cortical neurones, provides them with the ability to continually adjust their responsiveness as a function of their history of synaptic aid action potential activity. Such processes can optimise, in time, the input-output sеnsitivity of each individual neurone.
Other processes that change the input-output transfer function of a neurone depending upon its past activation are the synaptic processes of facilitation and depression. An important question is the expression of these processes in the variety of neuronal types which coexist in every microcolumn of the cerebral cortex, whatever the sensory, motor or cognitive specialisation of the cortical area.
Alex Thomson, who works at University College London, gave a state-of-the-art review of « Frequency and Pattern-Dependent properties of Local Neocortical Circuits ». She showed the diversity of cell types and synaptic connections in the cerebral cortex, and she reviewed the relations between the morphological diversity and the functional diversity at different classes of synaptic connections. Ву combining dual recordings of synaptically connected neurones, immuno-fluorescence and morphological reconstruction, her work revealed the richness of functional properties that can be found in the neocortex. She demonstrated that different synaptic pathways behave differently in response to repeated stimulation. Some show facilitation, some depression, and some show complex mixtures of the two. There was considerable discussion about the functional implications of such phenomena. In particular, the work of Alex Thomson suggests that different classes of neurones in the same microcolumn of the cortex generate different processing information flows, that can store information related to different patterns of coactivation, involving both increases and decreases in the gain of the input-output transfer functions.
The next level of organisation to understand is the interaction between different information flows generated by the connective architecture of the cerebral cortex, both within and between cortical areas. Yves Frégnac (CNRS, Gif sur Yvette, France) presented a talk on the « Dуnamics of interaction between feedforward, recurrent and lateral inputs during visual cortical processing: intracellular synaptic physiology and psyсhophysical correlates ». Hе started by noting that, much work on the role of thalamocortical pathways in sensory processing, is dominated by mechanisms involving feedforward connectivity. Thus, in the case of the mammalian visual system, it is well established that visual receptive fields are topographically organised in ON and OFF discharge zones, and the genesis of orientation selectivitу at the cortical level results from the string imprint of the feedforward input. However, Fregnac then went on to present a range of experiments showing that visual cortical cells have access to a much wider zone of the visual field than expected from the precision of feedforward projections. In particular, he argued that more time than previously thought is available for the primаry visual cortical network to process incoming input through recurrent and lateral pathways, before information is relayed along the chain of cortical processing. The arguments are based on intracellular recordings, patch and sharp, in the primary visual cortex. lis findings suggest that subthreshold responsesin the unresponsive region surrounding the classical discharge field, result from the integration of visual activation waves spread by slowly conducting horizontal axons within primary visual cortex. Furthermore, there is a dеlaу of several tens of ms between the earliest signs of cortical subthreshold activation (when the neurone first starts receiving the input) and the mean latency of the output V1 signal (when the neurone sends an output message), probably due to a shunting DABA inhibition. One mауextrapolate that local inhibitorу circuits put the cortical cells « on hold », waiting for the confirmation or invalidation of the relevance of the input by the rest of the network. Based on neurophysiological and psychophysical results, Yves Fregnac proposes that neural activity spreading through facilitatory and inhibitory long range horizontal intracortical connections, modulates the response latencies from V1 neurons.
Jean Bullier, from the Centre de Recherche Cerveau & Cognition in Toulouse, showed results on « The role of top-down connections in visual information processing by the cortex ». He argued that, despite the large number of feedback connections Between visual areas of the mammalian brain, very few studies have been devoted to the elucidation of the role of these connections in cortical processing. Hе then reviewed recent work that has studied the function of feedback connections from area MT to areas V1, V2 and V3 in macaque monkeys. lis results indicate that feedback connections serve to amplify the differences between the responses to the figure versus the background in a figure/ background stimulus configuration. This effect is particularly evident for 1ow visibility stimuli that are masked bу noisy backgrounds, and suggests that feedback connections maу play an important role in scene segmentation. Hе also described other work that has looked at the timing of the effects of feedback connections. Given that higher order areas contain neurones whose responses tend to lag behind those of lower order areas, it is surprising to learn that most effects were observed from the earliest part of the visual response and even for neurones with short latencies to visual stimuli. These results suggest that feedback connections appear to be recruited vert’ early for the treatment of visual information, possibly through a fast M (magnocellular) pathwaу.
2- What is the functional role of the temporal variability observed in the neuronal signals during visual information processing?
A constant observation in recording neuronal activation during visual information processing in the diffеrеnt areas of the cortex is the large variability of this activation, that contrasts with the high acсurасy of the visual processing.
Alain Destexhe (who started his research career as a doctoral student in Agies Babloyantz’s laboratory in Brussels) gave a talk entitled « Dynamics of dendritic integration in pyramidal neurones during active states in the neocortex ». It is well known that neo-cortical pyramidal neurones in vivo are subject to an intense synaptic background activity, but little is known of how this aсtivity affects cellular responsiveness, or what function it may serve. By blocking the background input, Destexhe found that, contrary to the intuitive idea that neurons might encounter difficulties when processing information in highly fluctuating states, cortical neurones are more sensitive and can detect events that would normally be undetectable. Alain Destexhe developed a precise biophysiсal model of pyramidal neurones in the cortex. Не showed, by computer simulation of the input/output transfer function of a neurone, that background aсtivity has a considerable impact on this transfer, allowing subthreshold inputs to become detectable. Intense network activity therefore seems to set pyramidal neurones into a particular state of responsiveness, which may be related to arousal or attentional mechanisms.
More theoretical work was presented by Klaus Pawelzik, a physicist working in Bremen, Germany, whose talk was entitled « Noise as Signal for Neuronal Populations ». Не developed a theoretical model of the input/output transfer function of cortical neurones, in a network of connected pyramidal neurones. Не performed a mathematical analysis of the way in which two populations of neurones interact. Ву deriving the equations of his model, he argued in favour of the surprising idea that noise in the nervous system might actually constitute a way of transmitting information.
Remy Lestienne (from the Institut des Neurosciences in Paris) made a remarkable synthesis of the « Variability and synchronicity on the sensory side of the central nervous system ». Hе reviewed a wide range of experimental work on response variability and temporal coding, not оnlу in vision, but also in the somatosensory system, olfaction and audition. He showed that the nervous system generates a vert’ precise timing for sensory processing, such as in the auditory and somatosensory system. By contrast, in further steps of information processing there can be a very large amount of response variability and loss in timing precision with respect to the stimulus. He discussed, in particular, the role of attention in spike variability in higher-level processing areas of the cerebral cortex. His review brought new insights to the question of the spatio-temporal coding of sensory information, at the various steps of its processing. Timing seems to be very precise in those sensory systems ruled by external clicks (e.g. in the specialized nuclei of the auditory system), but much less precise in the cortical areas dealing with internally time-coordinated information.
Barry Richmond from NIH at Bethesda asked the same question of the functional role of the variability in neuronal activation. He gave a challenging talk entitled « Listening in on the brain: What is noise and what is alphabet? » He pointed out that despite the intensive effort that has gone into trying to define which aspects of neuronal responses carry information, especiаlly for higher brain functions such as perception and memory, the structure of the code used by single neurones is still not yet certain. Barry Richmond has used a combined experimental and theoretical approach for determining which aspects of neural activity are useful for encoding information. He showed that the aspect of neuronal responses that changes most obviously with experimental conditions is the spike count – that is the number of spikes per timе unit. It is not diffiсult to observe that the pattern of activity, at least as a pattern of rate variation (the peri-event time histogram), is also different across stimuli. If these were the only factors in the response, deciphering neuronal responses might have gone further than it has. However, under most conditions the numbers of spikes and their timings vary considerably across repetitions of seemingly similar experimental conditions. The origin of this variability has crucial implications for neuronal information transmission. Does it represent information hidden in the response, or is it really just variability? This latter issue is intensively debated, because of proposals that exact spike times might be critical for the emergence of higher functions. In that case, the variability would be a consequence of extra information being carried in the time sequences of the spikes.
3-What is the functional role of oscillations and synchronies observed in the neuronal signals during visual information processing?
It has been suggested from both experimental and theoretical results that oscillations of neuronal activity at different frequencies and synchrony of spike activity between different neurons in large populations can have an important role in integrating and/or memorizing sensory information. In this meeting we had the chance to compare the properties of these oscillations at three different levels of brain organisation: at the cellular level (Richard Traub), at the level of large populations of neurones in the visual cortex (Reinhard Eckhorn) and at the level of different cortical areas in the human brain (Catherine Tallon-Baudry). New hypotheses emerged from this comparison.
Roger Traub, from the School of Medicine at Birmingham University made a synthesis on the generation of oscillations at different frequencies and at the level of local cortical and thalamocortical networks isolated in vitro. Hе showed that it is possible to generate two different types of oscillations — in the alpha range (10 Hz) and gamma range (40 Hz) in vitro. Very interestingly, these two oscillations generate two states of collective and long-lasting activation in the neuronal populations, with two different effects on the neuronal input-output transfer function. The results at the neuronal level are consistent with the hypothesis that gamma range oscillations increase the functional connectivity in a selected population of neurons, and that this could be the substrate of binding. In contrast, oscillatory activity in the alpha range activity has the effect of maintaining the activation in the population and could, as a result, be involved in short term or working memory. Roger Traub presented new results showing that there exists in the brain another type of oscillation with even higher frequency: ultrafast (>100 Hz) neuronal oscillations. These vеry intriguing data suggest that gap junctions between axons of pyramidal neurones could be involved in the generation of these ultrafast neuronal oscillations. The functional role of these synchronized oscillations could be a stronger efficacy of the output activation of a local population of neurones.
The functional role of oscillations in the gamma range in large populations of neurones was explored by Reinhard Eckhorn from Marburg in Germany. Hе gave a talk entitled « Flexible Cortical Synchronization Suggests Neural Principles of Visual Feature Grouping ». Hе summarised a number of recent studies on the visual cortex of perceiving monkeys that refer to the binding-bysynсhronization hypothesis. The experimental observations on multiple neuronal recording in the visual cortex show that gamma synchronizatiоn (35-90 Hz) is related to processes of feature binding, figure-ground segregation, and determination of object continuity. From these observations, he presented a neural network mode! that can explain, by local inhibitory feedback and Hebbian plasticity, the generation of rhythmic activities. Hе could also explain, by distance-dependent delays, the restricted gamma synchronization range and travelling waves which maу be important to perceive object continuity and limits (figure-ground segregation). Hе argued that long-range interactions could be mediated by the amplitude envelopes of local gamma processes. The relevance of the latter concept was illustrated with intracortical recordings from monkеy and subdural recordings from human subjects performing cognitive tasks.
Further evidence that cortical oscillations could play an important role in perception came from Catherine Tallon-Bаudry, who works at INSERM U280 in Lyon. She gave a talk entitled « Gamma oscillations in humans ». She argued that oscillatory synchronization could provide a link between the different areas engaged into the same perceptual or cognitive process. Bуusing both surface and intracranial EEG recording in humans, she has reрeatedly observed transient high-frequency (30-60 Hz) activity, between 200 and 300 ms after stimulus onset, which reflects the synchronized activity of large neural populations. Long-lasting bursts of oscillatory aсtivity, reflecting within-area synchronization, are observed in the gamma and beta bands, and show distinct functional variations. Between-area synchronization is also observed. At recording sites separated by about 1 cm, synchronization was not task-dependent, whereas long-distance (> 1 cm, distinct functional areas) synchronization was highly sensitive to the task performed. Her results show that oscillations in large populations of neurons observed in human, have similar properties as in monkey brain, and that data obtained at the scalp level can be explained by underlying local and long-distance neural oscillatory synchrony. Catherine Tallon-Ваudry suggested that variations of oscillations frequencies and synchronies related to stimulus type, or with the task to be performed, are consistent with a role of these synchronies in three types of information processing: the bottom-up construction of a coherent percept, a top-down influence in a visual search task, and the rehearsal of a stimulus in short-term mеmory.
Theoretical insights in the role of oscillations and synchronies in memоry were provided by Jacques Demongeot, from Grenoble. Hе talked about « The role of positive regulation circuits in mnesic evocation ». Hе explained, on a theoretical basis, that the role of positive circuits in the connectivity matrix of a neural network appears crucial for mеmоry: the number of attractors of the corresponding dynamical system is related to the existence of such positive regulatory loops, as well as the possibility to recall these attractors by stimulating a system made of subnetworks having the same internal connectivity. Jacques Demongeot presented a simple model for the hippocampus, showing three main complementary properties, i.e. muiti-attractority, synchronizability and desynchronizability, the three being essential for the memory function (storage and evocation)
4- What is the neural basis of rapid image processing of natural scenes?
The Tact that the brain can recognize complex natural scene in a few hundred milliseconds using neural components that are slow (by comparison with the digital components in a computer), is a major challenge for theories of brain function.
Simon Thorpe, from the Brain and Cognition Research Centre in Toulouse, reviewed a number of studies that have investigated the speed of computation in the brain. He gave a challenging talk on « Timing constraints and neural computation ». In the early 80s, Jerry Feldman’s so-called 100-step rule was one of the major factors that stimulated the development of connectionist and PDP models of brain function. Today, these timing constraints have become even more severe. For example, in the case of the human visual system, the decision as to whether a previously unseen photograph contains a target category (for example, animal or means of transport) can be made on the basis of only 150 ms of processing. In monkeys, the available time is probably only 100 ms. Such constraints imply that a considerable amount of processing can be achieved on the basis of rapid feed-forward mechanisms and imply that conventional coding schemes mаy be simply incompatible.
Rufin van Rullen, who did his doctorate with Simon Thorpe in Toulouse and is currently working as a post-doc in Christof Koch’s lab at CalTech, gave a talk entitled « Neural Mechanisms for UltraRapid Visual Processing: Theoretical Issues ». First he described the studies using Event-related potentials, that show the visual system can achieve a high-level representation of the visual input in about 150 ms (other processes can be observed earlier, but these mostly correspond to low-level sensоry analysis). Such findings constitute string theoretical constraints on the underlying neural mechanisms. Specifically, it would seem to imply that, the time available for information transmission between two successive processing stages, limits information transmission to roughly one spike per neurone, effectively ruling out conventional rate coding strategies. Other codes could encode information the spatio-temporal pattern of discharges in a `spike wave’ propagating through the system. Inc such code, `rank order coding’, was compared to other classical rate codes in the context of retino-geniculate information transmission, and proves highly effective. Van Rullen argued that, even when each neuron only generates one spike, Ultra-Rapid Visual Processing might rely on an essentially feed-forward architecture. The feed-forward propagation of a single spike wave, carrying information in its spatio-temporal structure, which is modified at each stage by (i) the neurones selectivities (ii) lateral interactions and (iii) top-down attentional influences, could be an ideal candidate mechanism to account for Ultra-Rapid Visual Processing.
Some very severe temporal constraints were provided by Christian Keysers who has been working in Professor David Perrett’s laboratory at the University of St. Andrews in Scotland. His talk, entitled « Minimal response duration for single cells in the macaque temporal cortex », attempted to answer questions such as « what is the minimal duration of a single cell response to an image? » and « what is the shortest time available for stimulus coding? » To answer such questions, theуrecorded from single neurones selective for complex patterns (e.g. faces) in the temporal cortex of the monkey while presenting images using rapid serial visual presentation. In this approach, images appear, one after the other, in a continuous sequence. These experiments demonstrate the visual system to respond to stimuli for ~60ms longer than their physical duration, an effect that provides precious extra time for stimulus coding. This neural persistence is surprising because it maу come at a cost: neuronal responses to a stimulus lasted for ~60ms into the time devoted to the processing of the next stimuli in the sequence, probably interfering with their processing. Such interference at this and prior processing stages maу be responsible for the decrease in peak firing and mutual information observed at high presentation rates. Keysers also reported psychophysical results that show an excellent correspondence between the response of single neurone in the temporal lobe and the performance of human subjects under similar conditions.
We now face the challenge of understanding the role of neurons of the temporal lobe in evеryday vision, with natural scenes that are vеry complex in terms of feature content. David Sheinberg, from Brown University in the U.S. gave a fascinating talk entitled « From perception to action: The role of temporal cortical neurons in noticing visual stimuli. » Неhas recorded the activity of these cells in the inferotemporal cortex (part of the temporal lobe) while monkeys perform tasks using more complex stimuli, such as ambiguous figures and natural scenes. Ву training the animals to reliably report what they see, Sheinberg and Logothetis have begun to tackle the problem of relating patterns of neural aсtivity in extended trials to the incoming stimulus, the observed response, and hypothetical internal states. They found that the activity of temporal cortical neurons, partiсularly as evidenced by patterns of short bursts, can be remarkably effective at signalling the presence of particular visual targets, even in the context of complex scenes. However, these signals only seem to occur just before the animal identifies the visual stimulus, bolstering the view that these cells participate in the process of « noticing » objects in the real world.
At the end of the workshop, Yves Burnod (INSERM 0483 Paris) organised a discussion session in which all the different themes of the meeting were dealt with. In the literature, there are several confrontations between apparently incompatible theories on the functional role of the temporal dynamics observed in neurones of the visual system during visual information processing. Another picture progressively emerged from the final discussion in « Les Treilles »: it was rather a consensus on the possible functional role of the different temporal aspects of information processing, observed at the different levels of brain organisation. The brain properties for visual information processing and perception appear to result from the cumulative effect of a set of dynamic processes at different levels of brain-organisation. These include intrinsic ionic channels properties; activity-dependent changes of synaptic efficacy in different types of cortical neurones; local interactions between types of connections (feedforward, lateral, feedback); interaction between background activity and stimulus-related signals, oscillations and synchronies of neuronal activation in different subsets of neurones at different temporal frequencies; encoding of information at different levels of time and precision (single spike or trains), both within neurones (synaptic conductances) and between neurones (spikes); dual information coding at the individual neuronal level and at the population level, parallel processing in a set of cortical areas.
It had been a fascinating meeting for all concerned, with a rich mix of experimental and theoretical work, at different levels of analysis: a consensus, not yet present in the literature, emerged between participants on the multiple temporal aspects of information processing in the brain.