Neural mechanisms of working memory in the prefrontal cortex
Shintaro Funahashi
Kokoro Research Center, Kyoto University, Kyoto, Japan
Working memory is a mechanism for short-term active maintenance of information as well as for processing maintained information. The dorsolateral prefrontal cortex (DLPFC) has been known to participate in working memory. The analysis of task-related DLPFC activity while monkeys performed a variety of working memory tasks revealed that delay-period activity is a neural correlate of a mechanism for temporary active maintenance of information, because this activity persisted throughout the delay period, showed selectivity to a particular visual feature, and was related to correct behavioral performances. On the other hand, information processing can be considered as a change of the information represented by a population of neurons during the progress of the trial. Using population vectors calculated by a population of task-related DLPFC activities, we demonstrated the temporal change of information represented by a population of DLPFC neurons during performances of spatial working memory tasks. Cross-correlation analysis using spike firings of simultaneously isolated pairs of neurons reveals widespread functional interactions among neighboring neurons, especially neurons having delay-period activity, and their dynamic modulation depending on the context of the trial. Functional interactions among neurons and their dynamic modulation could be a mechanism of information processing in the working memory processes.
_________________________________________________________________________________________
Can attractor network models account for the statistics of firing during persistent activity in prefrontal cortex?
Nicolas Brunel
Neurophysics and Physiology, Université Paris Descartes, and CNRS, Paris, France
Persistent activity observed in neurophysiological experiments in monkeys is thought to be the neuronal correlate of working memory. Over the last decade, network modelers have strived to reproduce the main features of these experiments. In particular, attractor network models have been proposed in which there is a coexistence between a non-selective attractor state with low background activity with selective attractor states in which sub-groups of neurons fire at rates which are higher (but not much higher) than background rates. A recent detailed statistical analysis of the data seems however to challenge such attractor models: the data indicates that firing during persistent activity is highly irregular (with an average CV larger than 1), while models predict a more regular firing process (CV smaller than 1). I will discuss how this feature can be reproduced in a network of excitatory leakly integrate-and-fire neurons.
_________________________________________________________________________________________
The computational limitations of balanced networks
C. van Vreeswijk
Neurophysics and Physiology, Université Paris Descartes, and CNRS, Paris, France and Franco-Israeli Laboratory of Neurophysics and System Neurophysiology
Computation in neural networks relies crucially on non-linearity. In neural networks in the balanced state the non-linearity of the neuronal transfer function becomes functionally unimportant. The disappearance of this non- linearity strongly limits the computational power of balanced networks. I will show in examples of balanced networks for associative memory how one can try to circumvent this limitation of balanced networks and discuss the problems with these solutions.
_________________________________________________________________________________________
Balanced spatial working memory
David Hansel
Neurophysics and Physiology, Université Paris Descartes, and CNRS, Paris , France and Franco-Israeli Laboratory of Neurophysics and System Neurophysiology
Neural activity persisting for several seconds is thought to be the neural correlate of working memory in cortex. It was found recently that during persistent activity spike trains are highly irregular, even more irregular than in spontaneous activity. We show that this apparently innocuous feature raises a fundamental difficulty if one holds that neuronal nonlinearities combined with recurrent excitation underly activity persistence as usually assumed. Instead, we argue that the key nonlinearities involved are synaptic and not neuronal. We assess this proposal in the framework of a network model representing a circuit in prefrontal cortex involved in spatial working memory. This lead us to suggest that short term plasticity recently discovered in synapses made by pyramidal cells in prefrontal cortex is crucial in spatial working memory.
_________________________________________________________________________________________
Active memory maintenance with short-term synaptic facilitatio
Gianliugi Mongillo*
Neurophysics and Physiology, Université Paris Descartes, and CNRS, Paris , France and Franco-Israeli Laboratory of Neurophysics and System Neurophysiology
Current theoretical framework holds that information is actively maintained in working memory through enhanced firing rates (delay activity). This would be achieved either via persistent activity reverberation within selective neural populations or as a result of intrinsic single-cell properties (i.e. bi-stability). Electrophysiological studies show, however, that delay activity increase can be modest, sometimes completely disappearing during part of the delay period. We therefore propose a new theoretical framework whereby working memory is sustained by calcium-mediated synaptic facilitation in the recurrent connections of neocortical networks. In this account, the presynaptic residual calcium is used as a 'buffer' which is loaded, refreshed and read-out by spiking activity. Due to the long time constants of calcium kinetics, the refresh rate can be very low, which results in a mechanism that is metabolically efficient and resistant to external interferences. The duration and stability of working memory can be effectively regulated by modulating the spontaneous activity in the network.
*joint work with: Omri Barak and Misha Tsodyks
_________________________________________________________________________________________
Working memory for saccadic eye movements in the parietal cortex
Shabtai Barash
Department of Neurobiology, Weizmann Institute of Science, Rehovot, Israel
Area LIP of the parietal cortex is related to saccades, to visual attention, and additional related cognitive processes, and contains neurons that show persistent activity in memory-guided saccades. The talk will focus on comparisons of the neuronal activity (1) during memory-saccades towards versus opposite the target's direction (prosaccades and antisaccades), and (2) during memory-saccades comprising unguided choice versus guided-choice. We will consider computational problems and other implications arising from the results.
_________________________________________________________________________________________
Low-dimensional network models for data from the prefrontal cortex
Christian Machens
Group for Neural Theory, Ecole normale supérieure3, Paris, France
During short-term memory maintenance, different neurons in prefrontal cortex (PFC), recorded under identical conditions, show a wide variety of temporal dynamics and response properties [1]. These data are a specific example of the more general finding that neural recordings from frontal cortices often reveal that different neurons have very different response characteristics. Modeling this complexity of responses has been difficult. Most commonly, some features of the responses are focused on, and models that fit those reduced features are built. But can the full complexity of responses be easily captured ? Here we attack the problem by fitting simple recurrent neural network models to the data.
Following the traditional approach, we first group neurons into different classes. When selecting neurons from a single class the estimation procedure yields a connectivity matrix with two populations of neurons coupled by mutual inhibition and self-excitation. The connectivity matrix has rank one and approximately agrees with a model we proposed earlier [2]. When selecting neurons from two classes, a connectivity matrix similar to that of the ring attractor network emerges, with a rank of two. The full complexity and richness of the observed neural dynamics, however, can only be captured when estimating a network architecture from the full set of neurons. In this case, the resulting connectivity matrix has rank five and its structure is dominated by randomness. Simulations of the resulting network reproduce the full data set. We show that several of the eigenvalues of the connectivity matrix are close zero, so that the network dynamics has either a constant or integrating flow along the respective dimensions. Finally, we discuss the consistency of the estimated connectivity matrices with the measured noise correlations. [1] Timing and Neural Encoding of Somatosensory Parametric Working Memory in Macaque Prefrontal Cortex. C.D. Brody, A. Hernandez, A. Zainos, and R. Romo, Cereb. Cortex 13:1196-1207, 2003. [2] Flexible control of mutual inhibition: a neural model of two- interval discrimination. C.K. Machens, R. Romo, and C.D. Brody, Science, 307:1121-1124, 2005.
_________________________________________________________________________________________
The computational principles and neural mechanisms underlying contraction bias
Yonatan Loewenstein
Depts. of Neurobiology and Cognitive Sciences and the Interdisciplinary Center for Neural Computation, Hebrew University, Jerusalem, 91904, Israel
It is well established that the estimated magnitude of memorized stimuli is biased: small magnitudes are overestimated and large magnitudes are underestimated, a phenomenon known as 'contraction bias.' In a previous study monkeys were trained to memorize the frequency of a vibrotactile stimulus (Base) and compare it with the frequency of a second stimulus (Comparison) while single unit activity was recorded in their prefrontal cortex (Romo et al. (1999), Nature, 399:470-473). We identified that the pattern of errors made by the monkeys is consistent with the contraction bias, providing an opportunity to study this phenomenon both at the level of behavior and at the level of neural activity. Here we address two questions: (1) What are the computational principles and (2) the neural mechanisms underlying the contraction bias?
(1) We show that contraction bias is consistent with Bayesian inference, in which a noisy measurement is combined with a-priori knowledge about the distribution of Base magnitudes in order to improve performance. According to the Bayesian hypothesis, increasing the level of uncertainty in the magnitude of the memorized stimulus enhances the bias. This uncertainty is a function of the delay between the Base and Comparison frequencies, as the performance level of the monkey decreases with the duration of the delay. Indeed, as expected from the Bayesian hypothesis, the longer the delay between the Base and Comparison frequencies, the greater the bias. According to the Bayesian hypothesis, monkeys utilize the prior distribution of Base frequencies in their decision making process. In order to study how the monkeys estimate this prior distribution, we analyzed the dependence of the monkeys' decisions on the recent history of stimuli presented to them in the experiment. We show that the estimated prior distribution depends mostly on the recent history of several previous trials. (2) The firing rate of many prefrontal cortex neurons during the delay period is a monotonic function of the Base stimulus frequency. It has been suggested that decisions in the discrimination task are made by comparing this activity with the neural representation of the Comparison frequency. Thus, the contraction of the memorized frequencies should be reflected in the activity of the prefrontal cortex neurons, resulting in the biased decisions. By studying how past trials affect neural activity in the prefrontal cortex, we seek to identify the neural correlate of the contraction bias.
_________________________________________________________________________________________
Mechanisms of mGluR mediated plateau potentials in entorhinal cortex neurons
Erik Fransen
Dept. of Computational Biology, Stockholm Brain Institute, School of Computer Science and Communication, Royal Institute of Technology, Sweden
Behavioral neurophysiological investigations have shown that principal cells in entorhinal cortex (EC) show spiking activities correlated with the different phases (including the delay and choice periods) of delayed matching tasks (Young et al 1997, Suzuki et al 1997). Human as well as animal studies have identified a working memory sensitivity to cholinergic blockade, as well as an involvement of metabotropic glutamate receptors (mGluRs). Moreover, persistent firing is believed to be a crucial mechanism for the delay period activity. We have shown that in synaptically isolated EC principal cells, repetitive application of an input can generate multiple persistent graded levels of firing (Egorov et al, Nature 2002, Fransen et al Neuron 2006). The process depends on the stimulation of metabotropic (muscarinic, mGluR) receptors, and the activation of a calcium sensitive nonspecific cationic current (CAN). These findings suggest that persistent firing could also be driven through synaptic activation of mGluRs. We have further studied the mGluR activation in related experimental work (Yoshida et al, submitted 2008). We found that, independent of muscarinic activation, and in the presence of ionotropic synaptic blockade, synaptic stimulation produced a long-lasting plateau potential. This potential is suppressed by group I mGluR antagonists, and enhanced by group I mGluR agonists. In this work, we have utilized computational modeling to elucidate possible mechanisms involved in this phenomena. The model has Hodgkin-Huxley type representations of ion channels and has a multi-compartmental representation of the dendrites. High-threshold calcium channels provide calcium to a local pool, modeled by 1-D diffusion, activating the CAN current. This current produces the slow plateau depolarization. Experimentally, muscarinic type 1 and mGluR1/5 activation converges on phospholipase C (PLC), and subsequent lipid messengers are believed to modulate the CAN channel. We have modeled this modulation as a shift in calcium sensitivity, analogous to the effects found for e.g. capsaisin and menthol on TRP-channels (Nilius et al 2005). Presently, we are studying how plateau amplitude depends on stimulation frequency, time and duration. Due to the slow dynamics of the cationic-induced plateau, synaptic input may be integrated over long time scales, and enable even low frequencies to participate in cell firing. More specifically, this may be an important property in working memory, where neurons need to maintain persistent activities over long time periods with low irregular firing rates.
_________________________________________________________________________________________
Mechanism for Top-down Control of Working Memory Capacity
Albert Compte
Institut d'Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS), Barcelona, Spain
Working memory relies on the activation of both prefrontal and parietal cortex, possibly with prefrontal cortex exerting top-down control. However, there is still no mechanistic description of either capacity limitations or top-down control of maintenance of information in working memory. Here, we propose that lateral inhibition in parietal cortex limits mnemonic capacity. However, at high loads, this inhibition can be counteracted by excitatory input from prefrontal cortex, thus boosting parietal capacity. We formulate this computationally in a biophysical cortical microcircuit model, and conceptualize it in a mathematical equation. Predictions from the model were confirmed in an fMRI study. The model provides a mechanistic framework for understanding top-down control of working memory, and specifies two different contributions of prefrontal and parietal cortex to working memory capacity.
_________________________________________________________________________________________
Short-term memory effects on visual perception
Valentina Anna Maria Daelli
SISSA- Cognitive Neuroscience, Via Beirut 2, I-34014, Trieste, Italy
Memory traces, stored in the form of attractors or appearing as the result of recent perceptual experience, can actively shape processing and categorization of visual stimuli. Electrophysiology in monkey IT cortex and computational modeling indicate the existence of categorical boundaries in the dynamics of cortical networks. Our psychophysical experiments in humans show that these boundaries can be shifted following recent visual experience, in adaptation and priming paradigms.
_________________________________________________________________________________________
Memory blends
Dov Sagi
Department of Neurobiology, Weizmann Institute of Science, Rehovot, Israel
How do perceptual memories interact? I will describe results from psychophysical experiments, using sequences of visual stimuli, showing that similarity between subsequent stimuli has a critical impact on the generated memory.
_________________________________________________________________________________________
Short term memory traces in neural networks
Surya Ganguli*
Sloan-Swartz Center for Theoretical Neurobiology, University of California, San Francisco
Critical cognitive phenomena such as planning and decision making rely on the ability of the brain to hold information in working memory. Many proposals exist for the maintenance of such memories in persistent activity that arises from stable fixed point attractors in the dynamics of recurrent neural networks. However such fixed points are incapable of storing temporal sequences of recent events. An alternate, and relatively less explored paradigm, is the storage of arbitrary temporal input sequences in the transient responses of a recurrent neural network. Such a paradigm raises a host of important questions. Are there any fundamental limits on the duration of such transient memory traces? How do these limits depend on the size of the network? What patterns of synaptic connections yield good performance on generic working memory tasks? To what extent do these traces degrade in the presence of noise? We use the theory of Fisher information to construct of novel measure of memory traces in neural networks. By combining Fisher information with dynamical systems theory, we find precise answers to the above questions for general linear neural networks. We prove that the temporal duration of a memory trace in any network is at most proportional to the number of neurons in the network. However, memory traces in generic recurrent networks have a short duration even when the number of neurons in the network is large. Networks that exhibit good working memory performance must have a (possibly hidden) feedforward architecture, such that the signal entering at the first layer is amplified as it propagates from one layer to the next. We prove that networks subject to a saturating nonlinearity, can achieve memory traces whose duration is proportional to the square root of the number of neurons. These networks have a feedforward architecture with divergent connectivity. By spreading excitation across many neurons in each layer, such networks achieve signal amplification without saturating single neurons.
*Joint work with Haim Sompolinsky and Dongsung Huh.
_________________________________________________________________________________________
The costs of visual working memory
Jochen Triesch
Johanna Quandt Research Professor, Frankfurt Institute for Advanced Studies
The capacity of visual working memory has been extensively characterized, but little work has investigated how occupying visual memory influences other aspects of cognition and perception. Here we show a novel effect: maintaining an item in visual working memory slows processing of similar visual stimuli during the maintenance period. Subjects judged the gender of computer rendered faces or the naturalness of body postures while maintaining different visual memory loads. We found that when stimuli of the same class (faces or bodies) were maintained in memory, perceptual judgments were slowed. Our results suggest there is interference between visual working memory and perception, caused by visual similarity between new perceptual input and items already encoded in memory.
_________________________________________________________________________________________
Mixed neuronal selectivity is important in recurrent neural networks implementing context dependent tasks
Mattia Rigotti*
Center for Theoretical Neuroscience, Columbia University, New York, NY, USA,Institute of Neuroinformatics, UNI/ETH Zurich, Switzerland, Department of Neurobiology and Kavli Institute for Neuroscience, Yale Univ. School of Medicine, New Haven, USA
Higher order animals show the remarkable ability to flexibly adapt their behavior according to the context. The execution of complex cognitive tasks can be modeled as a series of event driven transitions between mental states, each encoding a certain disposition to behavior or a specific sensori-motor decision. In this work, we hypothesize that these mental states are instantiated neuronally by recurrent circuit dynamics, in the form of stable attractors of the neural activity. We show that the mathematical conditions for the attractors and the event driven transitions can be satisfied only if neurons are selective to combinations of internal mental states and sensory stimuli. One possible way to generate such mixed selectivity is to introduce neurons whose afferent connections have random synaptic strengths. This approach has at least three highly desirable features. First, in spite of the combinatorial explosion of possible neurons with mixed selectivity, the number of needed randomly connected neurons grows only linearly with the number of relevant task events and contexts, which makes a reasonably sized network able to execute extremely complex cognitive tasks. Second, the firing patterns of neurons of the simulated proposed network, capture several aspects of the activity recorded in prefrontal cortex and other brain areas involved in a complex cognitive processes. The activity is self-sustaining in the absence of events, rule selective, and highly heterogeneous. Third, the introduction of randomly connected neurons accelerates the convergence of learning algorithms and it can be exploited to rapidly learn complex behavioral tasks. In conclusion we think that mixed selectivity, so widely observed in the living brain, can be an important and general functional principle for executing complex cognitive tasks.
*Joint work with Daniel Ben Dayan Rubin, Xiao-Jing Wang and Stefano Fusi
Unconscious determinants of free decisions in the human brain
_________________________________________________________________________________________
John-Dylan Haynes
Bernstein Center for Computational Neuroscience, Berlin, Germany
There has been a long debate whether subjectively "free" decisions are determined by brain activity ahead of time. Previous claims that subjective decisions are preceded by brain activity have been highly criticized as inaccuracies in the participants' subjective reports. Also, it has remained unclear whether an intention to act is initiated in motor-related brain regions, or if high-level brain areas are involved. Here we use a combination of statistical pattern recognition and fMRI to show that the outcome of decisions can be decoded from brain activity in prefrontal and parietal cortex even up to ten seconds before they enter awareness. This delay is too long to be accounted for by inaccuracies in measuring the onset of conscious intentions. Instead it presumably reflects the operation of a network of high-level control areas that operate at a slow timescale and begin to prepare an upcoming decision long before it enters awareness. This suggests that our free choices can be determined by brain activity much earlier than commonly appreciated.