Brain Dynamics on Multiple Scales - Paradigms, their Relations, and Integrated Approaches

The dynamics of complex networks has been the subject of investigation of a large body of research in the area of nonlinear physics (Strogatz, 2001). Dynamical units, coupled together through specific network structures give rise to complex phenomena that can be solved analytically and numerically. Particular complexity arises when the networks have non-trivial connection topologies (Watts and Strogatz, 1998) and time-delayed interactions (Yeung and Strogatz, 1999), which increase the dimensionality of the system and enrich the repertoire of dynamical behaviours that the network can exhibit. The brain, with its billions of interconnected neurons, falls in the far end of network complexity and can hardly be solved analytically. However, insights into puzzling empirical observations can be obtained using simplified network models exhibiting similar phenomena (Breakspear, 2017, Cabral et al., 2017). One of such puzzling cases is the spontaneous emergence of macro-scale rhythms spanning 2 orders of magnitude, between 1 and 100Hz. Almost a century after the first EEG recordings (Berger, 1924), and despite their evident implications for brain function, the generative mechanisms of such macro-scale brain rhythms, with periods up to 1 second, remain under debate. For a system to display oscillations with a given frequency, there needs to be a combination of time constants that define its periodicity. While the action potential of a single neuron lasts only a couple of milliseconds, time constants arising from excitatory and inhibitory synaptic mechanisms together with finite axonal propagation speed allow for slower rhythms to be generated at the neural network level. However, the time constants within a neural mass have only been shown to generate rhythms in the gamma frequency range (30-100Hz), both experimentally in-vitro (Buhl et al., 1998) and theoretically (Brunel and Wang, 2003). As such, all indicators suggest that rhythms slower than the gamma range implicate connectivity to other brain structures. However, wether there is a rhythmic external drive coming from the thalamus or from higher order cortex remains unclear. In the early nineties, Niebur, Schuster and Kammen (1991) studied the behaviour of coupled oscillators in the presence of time delays and reported ‘a novel form of frequency depression where the system decays to stable states, which oscillate at a delay- and interaction-dependent reduced collective frequency’. Following the findings of Niebur and colleagues, we investigate the behaviour of brain areas with an intrinsic frequency in the gamma-band (40Hz) when coupled together with realistic whole-brain connectivity and with long-range transmission delays in the order of ~10ms (Cabral et al., 2014, Cabral et al., in preparation). We find that the system displays a broad range of stable collective frequencies ranging between 1 and 100Hz, which can be nicely tuned by an interplay between the global coupling weight and the local oscillatory drive. The evidence of Connectome Frequencies in the same range of the empirically-observed macro-scale brain rhythms reveals that time-delays between brain areas are not only non-negligible but likely play a role in shaping the brain in dynamical networks with function-specific resonant frequencies. Strogatz SH (2001) Exploring complex networks. Nature 410:268-276. Watts DJ, Strogatz SH (1998) Collective dynamics of 'small-world' networks. Nature 393:440-442. Yeung MKS, Strogatz SH (1999) Time Delay in the Kuramoto Model of Coupled Oscillators. Physical review letters 82:648-651. Breakspear M (2017) Dynamic models of large-scale brain activity. Nature Neuroscience 20 (3), 340-352. Cabral J, Kringelbach M, Deco G (2017) Functional Connectivity dynamically evolves on multiple time-scales over a static Structural Connectome: Models and Mechanisms. NeuroImage. In Press Berger H; Gray, CM (1929). "Uber das Elektroenkephalogramm des Menschen". Arch Psychiat Nervenkr. 87: 527–570. Cabral J, Luckhoo H, Woolrich M, Joensson M, Mohseni H, Baker A, Kringelbach ML, Deco G (2014b) Exploring mechanisms of spontaneous functional connectivity in MEG: how delayed network interactions lead to structured amplitude envelopes of band-pass filtered oscillations. NeuroImage 90:423-435.

Recent research on brain functions using whole-brain imaging methods have led a radical change in the interpretation of the brain state during rest: rather than considering the brain inactive and simply reacting to incoming stimuli, the prevailing hypothesis now is that the brain is inherently active in an organized way at rest to be optimally prepared for stimulus processing. A vast amount of fMRI studies looking at BOLD fluctuations at rest showed correlated amplitude fluctuations in different brain areas that resembles networks activated during tasks. It has been proposed that these resting state networks reflect a sort of “constant inner state of exploration” to make the system optimally prepared for a given impending input and thus influencing perception and cognitive processing. While this idea intuitively makes sense, the fluctuations seen with fMRI are too slow to prepare for a given unpredictable input and to allow a fast and adequate reaction. In order to mediate complex mental activities and optimally respond to the rapidly changing information input, the networks have to reorganize in different spatial patterns on a sub-second time scale (Bressler, Brain Res Rev, 1995). MEG and EEG can record fluctuations on this time scale and are thus better suited to study the fast dynamics of resting states and their influence on stimulus processing. An increasingly recognized way to study the spatial and temporal properties of resting-state networks recorded with multichannel EEG is based on the concept of EEG microstates, which are defined as global patterns of scalp potential topographies recorded with multichannel EEG arrays that dynamically vary over time in an organized manner. More concretely, the broad-band spontaneous EEG at rest can be described by a limited number of scalp potential topographies that remain stable for a certain period of time (80-120 ms) before rapidly transitioning to a different topography that remains stable again. These discrete epochs of topographic stability have been called “microstates” with the idea that the scalp potential field describes the momentary state of global neuronal activity and that changes of the topography of this field indicate changes of the global coordination of neuronal activity over time (Lehmann et al., Scholarpedia 2009). Several studies demonstrated that disturbances of mental processes related to neurological and psychiatric conditions manifest as changes in the temporal dynamics of specific microstates. Combined EEG-fMRI studies and EEG source imaging indicated that EEG microstates are closely related to resting state networks observed with fMRI. The scale-free properties of EEG microstates time series explain why similar networks can be observed at such different time scales. This talk will l give an overview of what brain electromagnetic microstates stand for, the available analysis procedures, the functional interpretations and their behavioral and clinical correlates known so far.

By introducing a ground-breaking technique known as Dynamic Causal Modeling (DCM), Karl Friston and his collaborators have provided a bridge between the macroscopic scale (e.g., whole brain EEG/MEG recordings) and the mesoscopic scale (local field potential generated by cortical columns). In this technique, a model of interconnected neural masses representing cortical columns is postulated as the generator of the brain activity associated with a particular task. This model is then inverted through Bayesian estimation so that mesoscopic connectivity parameters can be estimated through non-invasive neuroimaging. It is unlikely that establishing such a direct bridge between the macroscopic and the microscopic scale could be made. The massive over parametrization of a microscopic model from a very sparse neuroimaging recordings would make any attempts of inverse modeling a most certainly futile endeavour. Furthermore, the simulation of extensive microscopic scale models is too demanding from a computational point of view to be performed in loop for inverse modeling in the context of a daily clinical use. However, we propose that the mesoscopic scale can act as a stepping stone in integrating knowledge across spatial scales to improve clinical neuroimaging. The physiologically detailed simulation of the brain involved in the Blue Brain Project is an excellent platform for the integration and synthesis of neuroscientific knowledge in various paradigms (patch clamping, single cell sequencing, ion channel study, etc.). Such simulation platform could be leveraged to feed a priori information in detailed mesoscopic scale neural mass models though Bayesian techniques. These models, tuned from the microscopic scale, could then be used for neuroimaging using the Bayesian framework and DCM. This talk will explore the potential and limits of such a cross-scale integration scheme.

Using a range of electrophysiological methods, we have extensively tested the proposal that the alpha rhythm (8–14 Hz) reflects an active mechanism of functional inhibition: decreased alpha activity facilitates processing whereas increased alpha functions to suppress task-irrelevant and/or distracting input, with clear benefits for behavioural performance. Here, I will present evidence in support of this framework, from (1) human MEG studies, showing that alpha power modulations reflect the direction of spatial attention, a mechanism under top-down control, with behavioural consequences; (2) MEG as well as intracranial recordings in humans, showing inter- and intra-subject variability of alpha frequency over areas and tasks; and (3) intracranial LFP and spike recordings in non-human primates, showing that the alpha rhythm phasically modulates spike firing rate. Key in this line of work is the combination of several recording levels, allowing us to draw bridges between (population-level) brain oscillations and behaviour, as well as between brain oscillations and single cell activity.

Neural population equations such as Wilson-Cowan equations, neural mass or field models are widely used to study brain activity on a large scale such as EEG, MEG or fMRI activity. How these large-scale models are linked to the properties of single neurons is however unclear. Here we derive stochastic mean-field equations for several interacting populations at the {\em mesoscopic} scale starting from a {\em microscopic} model of interconnected generalized integrate-and-fire (GIF) neurons. Each population consists of 50 -- 2000 neurons of the same type but different populations account for different neuron types. On the microscopic level, the spiking of various cortical neuron types can be well predicted by the GIF model. On the mesoscopic scale, the stochastic integro-differential equations that we find account for both finite-size fluctuations of the population activity and pronounced spike-history effects in single-neuron activity such as refractoriness and adaptation. The mesoscopic dynamics reproduces the rich stochastic population dynamics obtained from microscopic simulations of the full spiking neural network model. Going beyond classical mean-field theories for $N->\infty$, our finite-$N$ theory describes nonlinear emergent dynamics such as finite-size-induced stochastic transitions in multistable networks and synchronization in balanced networks of excitatory and inhibitory neurons. We use the mesoscopic equations to rapidly integrate a model of a cortical microcircuit consisting of eight neuron types, which allows us to predict spontaneous population activities as well as evoked responses to thalamic input. Our theory establishes a general framework for modeling finite-size neural population dynamics based on single cell and synapse parameters and offers an efficient approach to analyzing cortical circuits and computations. Therefore, we expect that our novel mesoscopic population theory will be instrumental for understanding experimental data on information processing in the brain, and ultimately link microscopic and macroscopic activity patterns. References: T. Schwalger, M. Deger, and W. Gerstner. Towards a theory of cortical columns: From spiking neurons to interacting neural populations of finite size. ArXiv e-prints, November 2016.

Neuronal cultures provide a simple yet versatile experimental platform to monitor the behavior of a living neuronal network, and model it through Physics toolboxes such as dynamical systems or network theory. These in vitro networks are prepared by plating an ensemble of neurons over a substrate, which quickly connect to one another and shape within a week a de novo assembly with rich spontaneous activity. An important trait of neuronal cultures is that they allow the investigation of a number of neuronal network phenomena at a mesoscopic scale, i.e. in populations of few thousand neurons. In our laboratory, and using state-of-the-art fluorescence imaging technology, we can monitor all the neurons in fields of view on the order of a hundred mm2, and with both high temporal and spatial resolutions. Such a detailed monitoring allows to link single neuronal dynamics and connectivity traits to collective phenomena. Specifically, we investigate open questions in neuroscience and brain research such as the emergence of spontaneous activity, the importance of spatial embedding, the repertoire of activity patterns, or the resilience of neuronal networks to chemical or physical damage. The versatility and easy access of neuronal cultures, as well as the ability to act on them, has made them very attractive as a complementary tool to brain research. This has become particularly important in the last years in the context of neurological disorders. Here we will show how advances in ‘induced pluripotent stem cells’, for instance, has facilitated the preparation of neuronal cultures with specific diseases, their monitoring through development, the characterization of altered dynamics, and the potential impact of specific therapies to treat the disease or reduce network damage. These investigations can be seen as an excellent proxy to later target specific diseases in actual brains.

Spontaneous waves of spiking activity are observed in the retina during development. This activity is thought to play a central role in shaping the visual system and retinal circuitry. Waves first occur at early embryonic stages of development and gradually disappear upon maturation. This process involves several time and space scales from molecular level (neurotransmitters), to neuron, to neurons population in the retina. I will present a model describing these different scales, accurate enough to reproduce experiments and to predict experimental results. This model can also be studied by tools from dynamical systems theory and bifurcations analysis. In my talk I will present part of this analysis linking it to biophysics and experiments.

The cortical microcircuit, the network comprising a square millimeter of brain tissue, has been the subject of intense experimental and theoretical research. A full-scale model of the microcircuit at cellular and synaptic resolution [1] containing about 100,000 neurons and one billion local synapses exhibits fundamental properties of in vivo activity. Despite this success, the explanatory power of local models is limited as half of the synapses of each excitatory nerve cell have non-local origins. We therefore set out to construct a multi-scale spiking network model of all vision-related areas of macaque cortex that represents each area by a full-scale microcircuit with area-specific architecture. The layer- and population-resolved network connectivity integrates axonal tracing data from the CoCoMac database with recent quantitative tracing data, and is refined using dynamical constraints. This research program raises methodological as well as technological questions: Are simulations at this scale feasible with available computer hardware [2]? Are full-scale simulations necessary, or can models of appropriately downscaled density be studied instead [3]? And finally: How can dynamical constraints be built into a high-dimensional spiking network model [4]? The talk systematically addresses these questions and introduces the required technology before outlining the data integration process [5]. The simulation technology has been developed on the K computer in Kobe and JUQUEEN in Juelich and is incorporated in the NEST simulation code. Simulation results reveal a stable asynchronous irregular ground state with heterogeneous activity across areas, layers, and populations. Intrinsic time scales of spiking activity are increased in hierarchically higher areas, and functional connectivity shows a strong correspondence with that measured using fMRI. The model bridges the gap between local and large-scale accounts of cortex, and clarifies how the detailed connectivity of cortex shapes its dynamics on multiple scales. [1] Potjans TC and Diesmann M (2014) The Cell-Type Specific Cortical Microcircuit: Relating Structure and Activity in a Full-Scale Spiking Network Model. Cerebral Cortex 24(3):785--806 [2] Kunkel S, Schmidt M, Eppler JM, Plesser HE, Masumoto G, Igarashi J, Ishii S, Fukai T, Morrison A, Diesmann M, Helias M (2014) Spiking network simulation code for petascale computers. Front. Neuroinform 8:78 [3] Van Albada S, Helias M, Diesmann M (2015) Scalability of Asynchronous Networks Is Limited by One-to-One Mapping between Effective Connectivity and Correlations PLoS Comput Biol 11(9): e1004490 [4] Schuecker J, Schmidt M, van Albada SJ, Diesmann M, Helias M (2017) Fundamental Activity Constrains Lead to Specific Interpretations of the Connectome PLoS Comput Biol 13(2): e1005179 [5] Schmidt, M., Bakker, R., Shen, K., Bezgin, G., Hilgetag, C.-C., Diesmann, M., van Albada, S.J. (2016) Full-density multi-scale account of structure and dynamics of macaque visual cortex. arXiv:1511.09364.

Neural systems have been modeled at different scales. One can start from a huge number of coupled differential equations describing individual neurons in discrete spatial space (neuronal networks), or work with locally averaged and homogenized representation in continuous space (neural fields), or even do away with space completely and investigate lumped models (neural mass models). In the course of model simplification through the scaling hierarchy, it is naturally desired to retain the rich variety of dynamics that is the hallmark of neural systems. This talk will consider the range of dynamics and complex behavior in theoretical models and indicate their possible connections to experimental data.

What happens in our brains when we consciously see, hear, feel, and know? Any theory that purports to clarify these processes needs to characterize neural mechanisms whose interactions generate dynamical states with properties that mimic the parametric properties of the individual qualia that we consciously experience, notably the spatio-temporal patterning and dynamics of the neural representations that represent these qualia. In addition to mirroring properties of subjective reports of these qualia, these dynamical states should generate observable data about these experiences that are collected in psychological and non-invasive neurobiological experiments on humans, and are consistent with psychological, multiple-electrode neurophysiological data, and other types of neurobiological data that are collected from monkeys who experience the same stimulus conditions. This talk summarizes evidence that Adaptive Resonance Theory, or ART, is accomplishing this goal. ART is a cognitive and neural theory of how advanced brains autonomously learn to attend, recognize, and predict objects and events in a changing world. ART has predicted that “all conscious states are resonant states” as part of its specification of mechanistic links between processes of consciousness, learning, expectation, attention, resonance, and synchrony. It hereby provides functional and mechanistic explanations of data ranging from individual spikes and their synchronization to the dynamics of conscious perceptual, cognitive, and cognitive-emotional behaviors. ART has now reached sufficient maturity to begin classifying the brain resonances that support conscious experiences of seeing, hearing, feeling, and knowing. The talk will review six different types of resonances, where they occur in our brains, and why they occur there; how they interact when we feel and know about what we see and hear; and some of the many psychological and neurobiological data in normal individuals and clinical patients during conscious and unconscious experiences that have not been otherwise explained. The talk will also mention resonances that do not become conscious, and why, and why not all brain dynamics are resonant, in terms of the computationally complementary organization of cortical processing streams. Along the way, the talk illustrate workshop themes, including the interplay between structure, dynamics, and function; information processing and surprise; brain oscillations and synchronization; functions and mechanisms at multiple scales and levels; and the organization of large-scale cognitive brain systems. References Grossberg, S. (2017). Towards solving the hard problem of consciousness: The varieties of brain resonances and the conscious experiences that they support. Neural Networks, 87, 38–95. http://www.sciencedirect.com/science/article/pii/S0893608016301800 Franklin, D. F., and Grossberg, S. (2016). A neural model of normal and abnormal learning and memory consolidation: Adaptively timed conditioning, hippocampus, amnesia, neurotrophins, and consciousness. Cognitive, Affective, and Behavioral Neuroscience, DOI 10.3758/s13415-016-0463-y, http://rdcu.be/m8um. Grossberg, S. (2013). Adaptive Resonance Theory: How a brain learns to consciously attend, learn, and recognize a changing world. Neural Networks, 37, 1-47.

What does it feel like to be a bat? Is conscious experience of echolocation closer to that of vision or audition? Or, echolocation is non-conscious processing and it doesn't feel anything? This famous question of bats' experience, posed by a philosopher Thomas Nagel in 1974 (Nagel, 1974), clarifies the difficult nature of the mind-body problem. Why a particular sense, such as vision, has to feel like vision, but not like audition, is puzzling. This is especially so given that any conscious experience is supported by neuronal activity. Activity of a single neuron appears fairly uniform across modalities, and even similar to those for non-conscious processing. Without any explanation on why a particular sense has to feel as the way it does, researchers even cannot approach the question of the bats' experience. Is there any theory that gives us a hope for such explanation? Currently, probably none, except for one. Integrated Information Theory (IIT), proposed by Tononi in 2004 (Tononi, 2004) has a potential to offer a plausible explanation. IIT essentially claims that any system that is composed of causally interacting mechanisms can have conscious experience. And precisely how the system feels like is determined by the way the mechanisms influence each other in a holistic way. In this talk, I will give a brief explanation of the essence of IIT and provide initial empirical partial tests of the theory, proposing a potential scientific pathway to approach bats' conscious experience. If IIT, or its improved or related versions, is validated enough, it will gain credibility to accept its prediction on rough nature of bats' experience. If we can gain a sophisticated insight as to whether bats' experience is closer to vision or audition, it is already a tremendously big step in consciousness science, which is just a first yet critical one, possibly a similar level of the breakthrough in cosmology in precisely estimating the age of the universe. Biography Dr Tsuchiya was awarded a PhD at California Institute of Technology (Caltech) in 2006 and underwent postdoctoral training at Caltech until 2010. Receiving a PRESTO grant from Japan Science and Technology (JST) agency, Dr Tsuchiya returned to Japan in 2010. In Jan 2012, he joined the School of Psychological Sciences at Monash University as an Associate Professor. Since 2013, he is an ARC Future Fellow. His main research interest is to uncover the neuronal basis of consciousness. Specifically, he focuses on 1) the scope and limit of non-conscious processing, 2) the relationship between attention and consciousness, and 3) the neuronal correlates of consciousness by analysing the multi-channel neuronal recording obtained in animals and humans and 4) testing a theory of consciousness, in particular, integrated information theory of consciousness.)

Based on Hebb's classsical concept of cell assemblies and my own analysis of neural associative memories, I will introduce a model of associatively connected cortical areas which is in principle able to do any computation. I will use the example of understanding simple sentences to show how a concrete computatioal task can be programmed and performed in such a network in a psychologically plausible way.

We consider a deterministic system of repulsively coupled Kuramoto oscillators, which are exposed to a fixed distribution of natural frequencies. As expected for antagonistic couplings, the attractor space is quite rich. We identify rather long transients that it takes the deterministic trajectories to find their stationary orbits in the attractor space. The stationary orbits show a variety of different periods, which can be orders of magnitude longer than the periods of individual oscillators. The smaller the width of the distribution about the common natural frequency is, the longer are the emerging time scales on average. Among the long-period orbits we find self-similar temporal sequences of temporary patterns of phase-locked motion on different time scales. The ratio of time scales is determined by the ratio of widths of the distributions. We interpret the long-period orbits as heteroclinic orbits, connecting unstable limit cycles. The effects are particularly pronounced if we perturb about a situation, in which a self-organized Watanabe-Strogatz phenomenon is known to happen, going along with a continuum of attractors and a conserved quantity. We compare the phase space for different values of the bifurcation parameter, deeply in the multistable regime and closer to the critical point, where the monostable regime turns into the multistable one. As we know from earlier results, noise can kick the system from one metastable state to the next. In case of disorder via the natural frequencies the motion is fully deterministic, but also leads to phenomena like physical aging, as it was formerly observed in [1] under the action of noise. The mechanism for generating long time scales and stable heteroclinic orbits and the high metastability under the action of noise may be of relevance for the dynamics that is found in the resting activity of the brain.

EEG slow waves and spindles are fundamental features of brain activity during non-rapid eye movement (NREM) sleep, and their occurrence represents the cornerstone for the ‘gold-standard’ EEG-based approach for distinguishing states of vigilance. In the last decade, however, it has become clear that waking, NREM sleep and REM sleep are not all-or-none, mutually exclusive states. Instead, mixed states characterised by a co-existence of sleep- and wake-like patterns of brain activity, or NREM and REM sleep features, are the norm rather than the exception across a wide range of conditions in both animals and humans. Both in vivo and in vitro recordings have revealed that sleep-like patterns, characterised by slow waves reflect the fundamental organisation of network connectivity within small networks. Notably, both long-term and immediate preceding state history affects the dynamics of global and local EEG slow waves and neuronal OFF (silent) periods during NREM sleep. Interestingly, cortical OFF periods and slow waves are also typical during waking, especially after sleep deprivation, and our data suggest that even during highly active behaviours local cortical “sleep” can occur, specifically during repetitive movement. The spatio-temporal dynamics of another hallmark of brain activity – sleep spindles – remains poorly understood. Evidence suggests that sleep spindles require a large-scale circuitry, including both cortical and thalamic networks. However, intracranial recordings in patients with drug-resistant epilepsy revealed that spindles are local events, rarely occurring across large distributed networks. We performed multichannel LFP and MUA recordings from the somatosensory cortex (SCx) in mice during spontaneous sleep in the absence of any explicit novel experience or learning during preceding waking. Mapping the distribution of spindle events revealed a strikingly heterogeneous pattern of their occurrence. In most cases spindles were restricted to small circumscribed territories within the SCx, clustered towards its anterior boundaries. Typically they faded away and were no longer discernible at electrodes within a mere few hundred micrometres, and were often undetectable in the ‘global’ EEG. Unexpectedly, we observed an absence of theta (6-9 Hz) activity and a persistence of NREM-like activity, dominated by slow waves, during REM sleep in the LFP channels where NREM spindles were apparent, resulting in sleep manifested as a unitary state within specific local cortical areas. Our results suggest a possibility that the spatio-temporal dynamics of slow waves and spindles do not merely reflect preceding sleep-wake history or experience, but reflects a fundamental organisation of subcortical-cortical circuitry, which underlies the emergence of local and global brain states.

Despite increasing insights into the diverse functions of sleep, the mechanisms by which sleep improves, or, conversely, by which the lack of sleep impairs cognitive function and information processing in cortical networks, is largely still not understood. This may be result of a more general lack in the understanding of what determines normal cortical network function and information processing. Recent experimental work has provided evidence that cortical dynamics in areas with higher cortical hierarchy is governed by slow timescales in individual neuron activity (Honey et al., 2012 Neuron; Murray et al., 2014 Nature Neuroscience; Chaudhuri et al., 2015 Neuron). Specifically, it was found that neuron dynamics exhibits a slow autocorrelation decay in these areas which provides them with a temporal storage capacity over long periods up to the order of several hundreds of milliseconds, much longer than the timescales governing the action potential generation. These slow timescales have been argued to allow information to be temporally stored and, consequently, be processed in different cortical areas and over extended times — a functional advantage that can improve information integration and the signal-to-noise ratio in short-term memory or decision-making processes, and that is generally believed to be beneficial for brain function and cognition. The question, however, whether these essential timescales of cortical information processing are changed in a wake-time dependent manner which, in turn, could account for reduced network function, is still open. At a more conceptual level, it is not clear what enables neuronal networks to exhibit such long timescales in the first place. Here, we comprehensively study these cortical timescales during sleep deprivation in rodents, humans and with a computational neural network model. We conclusively show — at the individual neuron level in freely behaving rodents and in EEG in humans — that the precise timescales governing cortical dynamics are increasingly shortened during sleep deprivation. Conversely, sleep recovers slow cortex dynamics. In detailed analyses, we find strong evidence that these network-function-compromising changes in timescales over the course of extended wake cannot be explained by the previously reported "local sleep" episodes alone (Vyazovskiy et al., 2011 Nature). Instead, our findings indicate that they can be linked to slow progressive changes in the network connectivity as put forward in the "Sleep Homeostasis Hypothesis" (SHY, Tononi and Cirelli, 2003; DeVivo et al., 2017 Science). Our data analyses combined with computational modeling suggest that the slow timescales arise only when the network is in a balanced regime which is increasingly perturbed during extended wake. Specifically, the slow timescales arise as a consequence of network dynamics poised in the vicinity of a critical state, which is increasingly perturbed during sleep deprivation (Meisel et al., 2013 Journal of Neuroscience). Our observations at the individual neuron level and in EEG extend previous findings and provide a novel perspective into the decline of information processing during extended wake. Beyond the sleep/wake cycle, the connection between slow timescales and network excitability uncovered in our study may provide the adequate framework to link cognitive deficits to alterations in the excitation-inhibition ratio observed in many diseases. J. D. Murray et al., Nat. Neurosci. 17, 1661 (2014) C. J. Honey et al., Neuron 76, 423 (2012) R. Chaudhuri et al., Neuron 88, 419 (2015) V. V. Vyazovskiy et al., Nature 472, 443 (2011) G. Tononi, C. Cirelli, Brain Res Bull 62, 143 (2003) De Vivo et al., Science (2017) Meisel et al., Journal of Neuroscience (2013)

Slow oscillations (SO) are a stereotyped activity pattern pervasively expressed during slow-wave sleep and deep anaesthesia by the cerebral cortex of many species. SO in sensorial cortices are known to mirror early neuronal processing of environmental stimuli, and occur simultaneously in cell assemblies at different cortical depths and positions as a concerted multiscale activity. Here, I will present some recent advances in the understanding of the mechanistic organization of such multiscale phenomenon, by following the chain of activations and silencing of the different layers of the rat visual cortex under deep anaesthesia. I will show that Up states initiate in layer 6 spreading upward towards the cortical surface, and that layer 5 assemblies give rise to hysteresis loops like in flip-flop computational units. I will provide evidence that layer 6 activation is not originated in the first-order thalamus, but rather it is likely due to the cortical input from Up wavefronts traveling across the cortical surface, unravelling a hierarchy of cortical loops. Changes in this multiscale organization of slow-wave activity when the awake state is approached will be also discussed.

We discuss the interplay of three neuronal time scales in both autonomously active neural networks and in self-organized embodied robots: The neural dynamics occurring on time scales of 1-10 ms, pre-synaptic transient short-term plasticity evolving for time scales of 100-1000 ms, and Hebbian post-synaptic plasticity needing 10-100 seconds. Using self-limiting Hebbian plasticity rules derived from the stationarity principle for statistical learning [1,2], we show that fully adapting Networks of the order of 100 Neurons remain homeostatically stable for at least several hours. Homeostatic stability is enhanced both by network size and when incorporating, in addition, short-term synaptic plasticity. We also point out, that the time scale of short synaptic plasticity allows for smooth motor control. In this context we simulate robots within the LPZrobots physics simulation environment [3], finding that short-term synaptic plasticity generates self-organized regular and explorative chaotic locomotion. These behavioral patterns correspond to limit cycles and chaotic attractors, which arise in the sensory-motor loop. We conclude that short term synaptic plasticity operates on time scales important for a vast array of cognitive tasks and for motor control. [1] Echeveste, R., Eckmann, S., Gros, C. (2015): The Fisher information as a neural guiding principle for independent component analysis. Entropy, 17(6), 3838-3856. [2] Echeveste, R., Gros, C. (2014): Generating Functionals for Computational Intelligence: The Fisher Information as an Objective Function for Self-Limiting Hebbian Learning Rules. Frontiers in Robotics and AI, 1, 1. [3] Martin, L., Sándor, B., & Gros, C. (2016). Closed-loop robots driven by short-term synaptic plasticity: Emergent explorative vs. limit-cycle locomotion. Frontiers in Neurorobotics, 10.

We typically experience the subjective “stream of consciousness” during wakefulness and thus associate it strongly with perception and behavioral responsiveness. However, consciousness can also be experienced during states when both perception and responsiveness are suppressed, like REM sleep or ketamine sedation. Here we aimed to identify neuronal correlates of consciousness, and distinguish them from correlates of responsiveness, in spontaneous brain activity. Specifically, we analyzed co-activation patterns among brain areas, defined as correlations between local amplitudes of gamma-band activity. We hypothesized to observe intermediate levels of gamma-band coupling and a balance between positive and negative correlations during conscious states, and a departure from such a balanced regime during loss of consciousness. We analyzed data recorded invasively from 4 macaque monkeys sedated with different anesthetics (experiment 1) and from 10 epilepsy patients in different sleep stages (experiment 2). Therefore, we were able to compare a range of states - including responsive and conscious (wakefulness), unresponsive but conscious (ketamine sedation, REM sleep), and unresponsive and unconscious (propofol sedation, NREM sleep) - in two distinct species. Amplitude of gamma-band activity was hyper-correlated (an excess of positive correlations and fewer anti-correlations) globally across cortex during both unconscious states (NREM sleep and propofol anesthesia) but not during wakefulness, REM sleep, or ketamine sedation. This finding indicates that the subjective “stream of consciousness” is lost when brain activity becomes excessively correlated.

A major goal in theoretical neuroscience is to predict response properties of sensory neurons from first principles. An influential attempt has been the ‘efficient-coding’ hypothesis, which posits that sensory systems maximise information encoded about natural stimuli. While successful, efficient coding ascribes equal importance to all sensory information, despite empirical evidence that neural systems prioritise behaviourally relevant (& not just statistically likely) stimuli. Further, efficient coding is well-posed only with additional ad hoc assumptions about biological constraints (e.g. metabolic or wiring costs), reducing its generality. ‘Information bottleneck’ (IB) can, in principle, overcome these limitations. Here, one maximises information about a ‘task-relevant’ variable, constrained on the total information encoded about the stimulus. IB naturally captures the dependence of the neural code on the ‘task-relevance’ of incoming sensory signals, and its objective function does not require extra biological assumptions for the coding cost. Unfortunately, the IB problem is extremely difficult to solve in all but the most trivial cases (i.e. gaussian or discrete data), which has greatly impeded its use in neuroscience. We recently proposed a tractable variational IB algorithm (varIB) that generalises to non-gaussian stimuli and non-linear input-output relations. Our framework is a powerful tool for constructing top-down models of task-relevant efficient coding. In various limiting cases, the varIB framework replicates either traditional efficient or sparse coding results. In general however, the new theory goes far beyond classical results, predicting how encoded features vary with external noise, internal coding constraints, and task-demands. We demonstrate that non-linear (‘kernelized’) extensions of the framework capture high-level phenomena such as the ‘filling-in’ of an occluded visual region. Finally, we use the framework convolutionally, to resolve the conflict between efficient temporal coding (which favours stimulus decorrelation) and predictive coding (which favours extraction of slow features).

The primate visual system recognizes partially occluded objects in natural scenes without difficulty, and this capacity is yet to be matched in computer vision. The neural basis of such robust object recognition, however, is poorly understood. Recent experimental results from primate area V4, an intermediate stage in the shape processing pathway, suggests that feedback from higher cortical areas such as prefrontal cortex (PFC) may be important for maintaining robust shape selective signals. Here we implement predictive coding to investigate possible underlying mechanisms. Predictive coding interprets feedback signals as predictions made by the higher cortices about the sensory stimulus underlying the activity of the lower cortical area. Predictive coding, therefore, has been hypothesized to be a method of efficient coding. By applying predictive coding to a two-layer, hierarchical Bayesian model of V4 and its efferent PFC, we show that feedback predictions combined with feed! forward sensory inputs result in robust shape-selective responses under partial occlusion. Furthermore, the temporal dynamics of V4 and PFC neuronal responses in experiments are well-predicted by the two-step inference of the model, composed of initial computation based on feedforward signals and delayed feedback influences. Thus, our study suggests possible computational roles of feedback signals in the visual system and their potential significance in object recognition.

Independent of their genetic, biochemical, pyhsiological origins, behavioural symptoms of many brain disease (e.g. Epilepsy, Parkinson’s disease, anxiety) are causally associated with changes in the electrical activity of the brain. When the disease-related activity is restored to normal levels, symptoms of the disease are also alleviated. Similar evidence is accumulating for several other brain diseases. This suggests that a certain subset of brain diseases can be treated as diseases of brain dynamics. Moreover, when viewed from the perspective of network dynamics, similarities among behaviourally dissimilar brain diseases become apparent. I will discuss recent progress in modelling of the aberrant activity in Parkinson’s disease and anxiety. For both diseases, we investigated how low-level changes in the neuron excitability and synaptic connectivity lead high-level disease-specific dynamical changes in the basal ganglia and amygdala. This investigation revealed common network mechanisms shared by the two diseases. Finally, I will describe a closed-loop stimulation strategy that uses minimal stimulation to not only change the brain activity state to normal levels but also recovers the network computations lost due to the disease condition.

Synchronization of neural activity in the brain is involved in a variety of brain functions including perception, cognition, memory, and motor behavior. Excessively strong, weak, or otherwise improperly organized patterns of synchronous oscillatory activity appear to contribute to the generation of symptoms of different neurological and psychiatric diseases. However, neuronal synchrony is frequently not perfect, but rather exhibits intermittent dynamics. The same synchrony strength may be achieved with markedly different temporal patterns of activity. I will discuss methods to describe these phenomena and will present the application of this analysis to the neurophysiological data in healthy brain, Parkinson’s disease, and drug addiction disorders. I will finally discuss potential cellular mechanisms and functional advantages of some of the observed temporal patterning of neural synchrony.

Different measures of directional influence have been employed to infer effective connectivity in the brain. When the connectivity between two regions is such that one of them (the sender) strongly influences the other (the receiver), a positive phase lag is often expected. The assumption is that the time difference implicit in the relative phase reflects the transmission time of neuronal activity. However, Brovelli et al. (2004) observed that, in monkeys engaged in processing a cognitive task, a dominant directional influence from one area of sensorimotor cortex to another may be accompanied by either a negative or a positive time delay. Here we show that a model of two brain regions, coupled with a well-defined directional influence, displays similar features to those observed in the experimental data (reproducing delay times and coherence spectra). Our model is inspired by the theoretical framework of Anticipated Synchronization, which is a form of synchronization that occurs when a unidirectional influence is transmitted from a sender to a receiver, but the receiver leads the sender in time. This regime can be a stable solution of two dynamical systems coupled in a sender–receiver configuration when the receiver is subjected to a negative delayed self-feedback. Therefore, we suggest that the primate cortex could operate in a regime of Anticipated Synchronization as part of normal neurocognitive function if the self-feedback is mediated by inhibitory synapses. Furthermore, we show that this counterintuitive phenomenon can also occur at the neuronal level in a 3-neuron motif in the presence of an inhibitory loop.

Information transmission and processing is generated in a distributed way across biological and artifical systems ranging from neural and gene regulatory circuits to self-organized communication networks in engineering. Across science, information handling in biological systems is often referred to as 'distributed', yet, how information may be specically communicated and dynamically routed in such systems is not well understood. Here we present an advance on what 'distributed' information routing actually means and what it might condense to qualitatively and quantitatively: We develop a theory to predict patterns of information routing in networks of arbitrarily complex connectivity as a function of an underlying collective dynamical reference state. We uncover how local modications in individual unit properties, the structure of interaction networks and external inputs provide mechanisms to flexibly change information routing through the entire network. Intriguingly, local modifications in one network part may change information routing even between two other parts, pointing to remote control in distributed routing mechanisms. These results offer novel directions of future research on information routing across complex distibuted systems, with specific implications for analyzing and possibly designing neural circuit activity. This is work with Christoph Kirst and Demian Battaglia Reference: Kirst et al.(2016) Dynamic Information Routing in Complex Networks, Nature Comm. 7: 11061 (2016). http://dx.doi.org/10.1038/ncomms11061

Brain development involves many random processes that interact to determine the connectivity between neurons. Despite all these random components, different individuals are able to generate similar behaviour. What are the fundamental structural properties that are common between different networks and lead to proper network function? I n close collaboration with neurobiologists from the University of Bristol (Alan Roberts and Steve Soffe) we have developed a new computational method to define synaptic connections (in the form of a “connectome”) between all neurons in the hatchling Xenopus tadpole spinal cord. The connectome is generated by a “developmental” process where the growing axons intersect dendrites and create connections. The resulting network has around 1,500 neurons with around 100,000 connections, and its statistical properties are similar to experimental measurements. We have developed a functional network of spiking (Hodgkin-Huxley) neurons and used the generated connectome to produce a pattern of neural activity. Remarkably, the generated activity is very stable and corresponds with the typical pattern seen in vivo during fictive swimming (anti-phase oscillations between left and right sides of the body). We derive a probabilistic model of connectivity in the tadpole spinal cord that generalises a large number of partial connectomes generated by the “developmental” process. For each pair of neurons we calculate the probability of a directed connection. Thus, the probabilistic model is defined by a matrix of probabilities for independent Bernoulli random variables. Using this probability matrix, we can generate an adjacency matrix representing specific network connectivity and using this connectome, the functional model reliably produces a pattern of spiking activity corresponding to swimming. To make progress in understanding the interplay between structure and function, we use the probabilistic model to calculate multiple structural characteristics of the connectivity graph, such as degree distributions, heterogeneity and others. We calculate these characteristics using the probabilistic model, without considering particular realisations of network connections. Therefore, the structural characteristics reflect general properties of the network. We show that the tadpole connectome is rather homogeneous, suggesting that the highly connected nodes (hubs) are not essential for network functionality (unlike the C. elegans network). We use the structural characteristics to study neuronal dynamics in the complete network as well as in different sub-networks. For example, we find that the minimal network to generate swimming includes neurons of two types: commissural and descending interneurons.

We present a study that combines modeling of neuronal activity and networks derived from neuroimaging data in order to investigate how the structural organization of the human brain affects the temporal dynamics of interacting brain areas. The dynamics of the neuronal activity is modeled with FitzHugh-Nagumo oscillators and the blood-oxygen-level-dependent (BOLD) time series is inferred via the Balloon-Windkessel hemodynamic model. The simulations are based on anatomical probability maps between considered brain regions of interest. These maps were derived from diffusion-weighted magnetic resonance imaging measurements. In addition, the length of the fiber tracks allows for inference of coupling delays due to finite signal propagation velocities. We aim to investigate (i) graph-theoretical properties of the network topology derived from neuroimaging data and (ii) how randomization of structural connections influences the dynamics of neuronal activity. The network characteristics of the structural connectivity data are compared to density-matched Erd\H{o}s-R\'{e}nyi random graphs. Furthermore, the neuronal and BOLD activity are modeled on both empirical and random (Erd\H{o}s-R\'{e}nyi type) graphs. The simulated temporal dynamics on both graphs are compared statistically to capture whether the spatial organization of these network affects the modeled time series. Our results support previous findings that key topological network properties such as small-worldness of our neuroimaging data are distinguishable from random networks. We also show that simulated BOLD activity is affected by the underlying network topology and the strength of connections between the network nodes. The difference of the modeled temporal dynamics of brain networks from the dynamics on randomized graphs suggests that anatomical connections in the human brain together with dynamical self-organization are crucial for the temporal evolution of the resting-state activity. Reference: S.Bayrak, V. Vuksanović, P. Hövel: Modeling functional connectivity on empirical and randomized structural brain networks, Differential Equations and Dynamical Systems (2017) in print

The brain is a complex system of neuronal populations, which establish functional connections during task execution and at rest. Whereas with BOLD fMRI “resting state networks” have been studied extensively, allowing to investigate the spatial properties of these functional networks, it cannot reveal fast interactions which are needed to explain sub-second neural processing. Recent studies have shown that Hidden Markov Modelling (HMM) of MEG data can reveal consistent patterns of brain states fluctuating at small time scales. We have extended this HMM framework for EEG, revealing that it is possible to extract a rich collection of such fast transient networks. Moreover, we investigate the BOLD correlates of these networks using simultaneously recorded fMRI data, validating that these networks correspond to the traditionally known networks.

A synergistic conceptualization of development is to view an individual’s development across the lifespan in terms of the development of self-regulated adaptive neurocognitive processes that are in constant exchanges with contextual influences. Adaptive, goal-directed behavior requires self-regulated action and behavioral control under specific environmental/task demands, for which smooth operations are implemented through dynamic interactions between cortical monitoring, hippocampal memory, and subcortical motivational processes. Through its widespread projections, the dopamine systems innervate frontal-hippocampal-striatal circuitry, which could serve as the regulating interfaces between cognition and motivation. This presentation will give a brief overview about lifespan development of dopamine modulation, followed by selected empirical findings to highlight prenatal perturbation, sensory-perceptual conflict, cognitive training, and incentives as different types of contextual influences that may interact with neuromodulation in affecting development- and/or aging-related differences in adaptive cognition and behavior.

Brain network is dynamic, i.e. fluctuates with time. Yet most studies on brain network have focused on characterizing the brain network as static. Only recent studies suggest that the reconfiguration of brain networks across time is crucially linked with complex cognition. Here, we investigated the relationships between the flexibility of brain network with flexibility of cognitive processing in young children. Resting state MEG data were recorded from 88 children, 3-4 years old and their performance on several cognitive tests. We estimated the flexibility of brain network by edge variability, a simple first derivative based measure. We observed children scoring high on cognitive flexibility showed higher brain network flexibility. Further such dynamic measures of network flexibility offered better classification performance than static measures of brain network. These results altogether suggest that the way the brain network dynamically reconfigure itself is directly linked with the flexibility of cognitive processing at an early stage of development.