Mathematical formulations of Hebbian learning

 Several formulations of correlation-based Hebbian learning are reviewed. On the presynaptic side, activity is described either by a firing rate or by presynaptic spike arrival. The state of the postsynaptic neuron can be described by its membrane potential, its firing rate, or the timing of backpropagating action potentials (BPAPs). It is shown that all of the above formulations can be derived from the point of view of an expansion. In the absence of BPAPs, it is natural to correlate presynaptic spikes with the postsynaptic membrane potential. Time windows of spike-time-dependent plasticity arise naturally if the timing of postsynaptic spikes is available at the site of the synapse, as is the case in the presence of BPAPs. With an appropriate choice of parameters, Hebbian synaptic plasticity has intrinsic normalization properties that stabilizes postsynaptic firing rates and leads to subtractive weight normalization. Received: 1 February 2002 / Accepted: 28 March 2002 Correspondence to: W. Gerstner (e-mail: wulfram.gerstner@epfl.ch, Tel.: +41-21-6936713, Fax: +41-21-6935263)     

Influence of temporal correlation of synaptic input on the rate and variability of firing in neurons

The spike trains that transmit information between neurons are stochastic. We used the theory of random point processes and simulation methods to investigate the influence of temporal correlation of synaptic input current on firing statistics. The theory accounts for two sources for temporal correlation: synchrony between spikes in presynaptic input trains and the unitary synaptic current time course. Simulations show that slow temporal correlation of synaptic input leads to high variability in firing. In a leaky integrate-and-fire neuron model with spike afterhyperpolarization the theory accurately predicts the firing rate when the spike threshold is higher than two standard deviations of the membrane potential fluctuations. For lower thresholds the spike afterhyperpolarization reduces the firing rate below the theory's predicted level when the synaptic correlation decays rapidly. If the synaptic correlation decays slower than the spike afterhyperpolarization, spike bursts can occur during single broad peaks of input fluctuations, increasing the firing rate over the prediction. Spike bursts lead to a coefficient of variation for the interspike intervals that can exceed one, suggesting an explanation of high coefficient of variation for interspike intervals observed in vivo.     

The impact of short term synaptic depression and stochastic vesicle dynamics on neuronal variability

Neuronal variability plays a central role in neural coding and impacts the dynamics of neuronal networks. Unreliability of synaptic transmission is a major source of neural variability: synaptic neurotransmitter vesicles are released probabilistically in response to presynaptic action potentials and are recovered stochastically in time. The dynamics of this process of vesicle release and recovery interacts with variability in the arrival times of presynaptic spikes to shape the variability of the postsynaptic response. We use continuous time Markov chain methods to analyze a model of short term synaptic depression with stochastic vesicle dynamics coupled with three different models of presynaptic spiking: one model in which the timing of presynaptic action potentials are modeled as a Poisson process, one in which action potentials occur more regularly than a Poisson process (sub-Poisson) and one in which action potentials occur more irregularly (super-Poisson). We use this analysis to investigate how variability in a presynaptic spike train is transformed by short term depression and stochastic vesicle dynamics to determine the variability of the postsynaptic response. We find that sub-Poisson presynaptic spiking increases the average rate at which vesicles are released, that the number of vesicles released over a time window is more variable for smaller time windows than larger time windows and that fast presynaptic spiking gives rise to Poisson-like variability of the postsynaptic response even when presynaptic spike times are non-Poisson. Our results complement and extend previously reported theoretical results and provide possible explanations for some trends observed in recorded data.     

Identifying and Tracking Simulated Synaptic Inputs from Neuronal Firing: Insights from In Vitro Experiments

Accurately describing synaptic interactions between neurons and how interactions change over time are key challenges for systems neuroscience. Although intracellular electrophysiology is a powerful tool for studying synaptic integration and plasticity, it is limited by the small number of neurons that can be recorded simultaneously in vitro and by the technical difficulty of intracellular recording in vivo. One way around these difficulties may be to use large-scale extracellular recording of spike trains and apply statistical methods to model and infer functional connections between neurons. These techniques have the potential to reveal large-scale connectivity structure based on the spike timing alone. However, the interpretation of functional connectivity is often approximate, since only a small fraction of presynaptic inputs are typically observed. Here we use in vitro current injection in layer 2/3 pyramidal neurons to validate methods for inferring functional connectivity in a setting where input to the neuron is controlled. In experiments with partially-defined input, we inject a single simulated input with known amplitude on a background of fluctuating noise. In a fully-defined input paradigm, we then control the synaptic weights and timing of many simulated presynaptic neurons. By analyzing the firing of neurons in response to these artificial inputs, we ask 1) How does functional connectivity inferred from spikes relate to simulated synaptic input? and 2) What are the limitations of connectivity inference? We find that individual current-based synaptic inputs are detectable over a broad range of amplitudes and conditions. Detectability depends on input amplitude and output firing rate, and excitatory inputs are detected more readily than inhibitory. Moreover, as we model increasing numbers of presynaptic inputs, we are able to estimate connection strengths more accurately and detect the presence of connections more quickly. These results illustrate the possibilities and outline the limits of inferring synaptic input from spikes.     

Pre & postsynaptic tuning of action potential timing by spontaneous GABAergic activity

Frequency and timing of action potential discharge are key elements for coding and transfer of information between neurons. The nature and location of the synaptic contacts, the biophysical parameters of the receptor-operated channels and their kinetics of activation are major determinants of the firing behaviour of each individual neuron. Ultimately the intrinsic excitability of each neuron determines the input-output function. Here we evaluate the influence of spontaneous GABAergic synaptic activity on the timing of action potentials in Layer 2/3 pyramidal neurones in acute brain slices from the somatosensory cortex of young rats. Somatic dynamic current injection to mimic synaptic input events was employed, together with a simple computational model that reproduce subthreshold membrane properties. Besides the well-documented control of neuronal excitability, spontaneous background GABAergic activity has a major detrimental effect on spike timing. In fact, GABA(A) receptors tune the relationship between the excitability and fidelity of pyramidal neurons via a postsynaptic (the reversal potential for GABA(A) activity) and a presynaptic (the frequency of spontaneous activity) mechanism. GABAergic activity can decrease or increase the excitability of pyramidal neurones, depending on the difference between the reversal potential for GABA(A) receptors and the threshold for action potential. In contrast, spike time jitter can only be increased proportionally to the difference between these two membrane potentials. Changes in excitability by background GABAergic activity can therefore only be associated with deterioration of the reliability of spike timing.     

Relationship between applicability of current-based synapses and uniformity of firing patterns

The purpose of this paper is to identify situations in neural network modeling where current-based synapses are applicable. The applicability of current-based synapse model for studying post-transient behavior of neural networks is discussed in terms of average synaptic current strength induced by per spike during one firing cycle of a neuron (or briefly per spike synaptic current strength). It was found that current-based synapse models are applicable in both situations where both the interspike intervals of the neurons and the distribution of firing times of the neurons are uniform, and where the firing of all neurons is synchronized. If neither the interspike intervals nor the distribution of firing times of the neurons is uniform or the reversal potential is between the rest and threshold potentials, current-based synapse models may be oversimplified.     

Integrative Synaptic Mechanisms in the Caudal Ganglion of the Crayfish

A study of activity recorded with intracellular micropipettes was undertaken in the caudal abdominal ganglion of the crayfish in order to gain information about central fiber to fiber synaptic mechanisms. This synaptic system has well developed integrative properties. Excitatory post-synaptic potentials can be graded, and synaptic potentials from different inputs can sum to initiate spike discharge. In most impaled units, the spike discharge fails to destroy the synaptic potential, thereby allowing sustained depolarization and multiple spike discharge following single pulse stimulation to an afferent input. Some units had characteristics which suggest a graded threshold for spike generation along the post-synaptic fiber membrane. Other impaled units responded to afferent stimulation with spike discharges of two distinct amplitudes. The smaller or "abortive" spikes in such units may represent non-invading activity in branches of the post-synaptic axon. On a few occasions one afferent input was shown to inhibit the spike discharge initiated by another presynaptic input.     

Spike timing and reliability in cortical pyramidal neurons: effects of EPSC kinetics, input synchronization and background noise on spike timing

In vivo studies have shown that neurons in the neocortex can generate action potentials at high temporal precision. The mechanisms controlling timing and reliability of action potential generation in neocortical neurons, however, are still poorly understood. Here we investigated the temporal precision and reliability of spike firing in cortical layer V pyramidal cells at near-threshold membrane potentials. Timing and reliability of spike responses were a function of EPSC kinetics, temporal jitter of population excitatory inputs, and of background synaptic noise. We used somatic current injection to mimic population synaptic input events and measured spike probability and spike time precision (STP), the latter defined as the time window (Deltat) holding 80% of response spikes. EPSC rise and decay times were varied over the known physiological spectrum. At spike threshold level, EPSC decay time had a stronger influence on STP than rise time. Generally, STP was highest (6 ms) triggered spikes at lower temporal precision (>or=6.58 ms). We found an overall linear relationship between STP and spike delay. The difference in STP between fast and slow compound EPSCs could be reduced by incrementing the amplitude of slow compound EPSCs. The introduction of a temporal jitter to compound EPSCs had a comparatively small effect on STP, with a tenfold increase in jitter resulting in only a five fold decrease in STP. In the presence of simulated synaptic background activity, precisely timed spikes could still be induced by fast EPSCs, but not by slow EPSCs.     

Integration of subthreshold and suprathreshold excitatory barrages along the somatodendritic axis of pyramidal neurons

Neurons integrate inputs arriving in different cellular compartments to produce action potentials that are transmitted to other neurons. Because of the voltage- and time-dependent conductances in the dendrites and soma, summation of synaptic inputs is complex. To examine summation of membrane potentials and firing rates, we performed whole-cell recordings from layer 5 cortical pyramidal neurons in acute slices of the rat's somatosensory cortex. We delivered subthreshold and suprathreshold stimuli at the soma and several sites on the apical dendrite, and injected inputs that mimic synaptic barrages at individual or distributed sites. We found that summation of subthreshold potentials differed from that of firing rates. Subthreshold summation was linear when barrages were small but became supralinear as barrages increased. When neurons were discharging repetitively the rules were more diverse. At the soma and proximal apical dendrite summation of the evoked firing rates was predominantly sublinear whereas in the distal dendrite summation ranged from supralinear to sublinear. In addition, the integration of inputs delivered at a single location differed from that of distributed inputs only for suprathreshold responses. These results indicate that convergent inputs onto the apical dendrite and soma do not simply summate linearly, as suggested previously, and that distinct presynaptic afferents that target specific sites on the dendritic tree may perform unique sets of computations.     

Generation of synaptic noise: selective involvement of neuronal subsets.

All central neurons are subjected to continuous and random variations of their membrane potential because of "spontaneous" activity in their presynaptic afferents. This activity, which is called synaptic noise, is presumed to be responsible for the uncertainty of the input-output relation in these cells. In the Mauthner cell of teleosts, noise is mainly inhibitory, and is generated by the release of neurotransmitter in a probabilistic manner. This inhibitory activity has been studied in detail previously. Taking advantage of this understanding, we have constructed a model of the inhibitory networks and their target in order to determine the conditions required to reproduce the main stochastic aspects of synaptic noise. We have used a combination of computer simulations and simple semianalytical arguments. We conclude that, surprisingly, cells in the presynaptic networks do not contribute equally to these background fluctuations. Rather, noise is generated primarily by the operation of subsets of afferent cells: the spectrum is either dominated by signals originating from interneurons which make few terminals on the Mauthner cell, or by the output of "burster" cells firing spike trains rather than single spikes. Both possibilities lead to specific predictions, one of which has already been verified.     

Consistency and Diversity of Spike Dynamics in the Neurons of Bed Nucleus of Stria Terminalis of the Rat: A Dynamic Clamp Study

Neurons display a high degree of variability and diversity in the expression and regulation of their voltage-dependent ionic channels. Under low level of synaptic background a number of physiologically distinct cell types can be identified in most brain areas that display different responses to standard forms of intracellular current stimulation. Nevertheless, it is not well understood how biophysically different neurons process synaptic inputs in natural conditions, i.e., when experiencing intense synaptic bombardment in vivo. While distinct cell types might process synaptic inputs into different patterns of action potentials representing specific “motifs” of network activity, standard methods of electrophysiology are not well suited to resolve such questions. In the current paper we performed dynamic clamp experiments with simulated synaptic inputs that were presented to three types of neurons in the juxtacapsular bed nucleus of stria terminalis (jcBNST) of the rat. Our analysis on the temporal structure of firing showed that the three types of jcBNST neurons did not produce qualitatively different spike responses under identical patterns of input. However, we observed consistent, cell type dependent variations in the fine structure of firing, at the level of single spikes. At the millisecond resolution structure of firing we found high degree of diversity across the entire spectrum of neurons irrespective of their type. Additionally, we identified a new cell type with intrinsic oscillatory properties that produced a rhythmic and regular firing under synaptic stimulation that distinguishes it from the previously described jcBNST cell types. Our findings suggest a sophisticated, cell type dependent regulation of spike dynamics of neurons when experiencing a complex synaptic background. The high degree of their dynamical diversity has implications to their cooperative dynamics and synchronization.     

Factors affecting slow regular firing in the suprachiasmatic nucleus in vitro

In isolated slices of hypothalamus, suprachiasmatic nucleus (SCN) neurons were recorded intracellularly. Blockade of Ca++ channels increased spike duration, eliminating an early component of the afterhyperpolarization (AHP) that followed evoked spikes. The duration and reversal potential of AHPs were, however, unaffected, suggesting that only an early, fast component of the AHP was Ca(++)-dependent. Unlike other central neurons that exhibit pacemaker activity, therefore, SCN neurons do not display a pronounced, long-lasting Ca(++)-dependent AHP. Extracellular Ba++ and intracellular Cs+ both revealed slow depolarizing potentials evoked either by depolarizing current injection, or by repolarization following large hyperpolarizations. They had different effects on the shape of spikes and the AHPs that followed them, however. Cs+, which blocks almost all K+ channels, dramatically reduced resting potential, greatly increased spike duration (to tens of milliseconds), and blocked AHPs completely. In contrast, Ba++ had little effect on resting potential and produced only a small increase in spike duration, depressing an early Ca(++)-dependent component and a later Ca(++)-independent component of the AHP. The relatively weak pacemaker activity of SCN neurons appears to involve voltage-dependent activation of at least one slowly inactivating inward current, which brings the cells to firing threshold and maintains tonic firing; both Ca(++)-dependent and Ca(++)-independent K+ channels, which repolarize cells after spikes and maintain interspike intervals; and Ca++ channels, which contribute to activation of Ca(++)-activated K+ currents and may also contribute to slow depolarizing potentials. In the absence of powerful synaptic inputs, SCN neurons express a pacemaker activity that is sufficient to maintain an impressively regular firing pattern. Slow, repetitive activation of optic input, however, increases local circuit activity to such an extent that the normal pacemaker potentials are overridden and firing patterns are altered. Since SCN neurons are very small and have large input resistances, they are particularly susceptible to synaptic input.     

Synaptic potential in the motor giant axon of the crayfish

Some electrical properties of the synapses between central giant axons (presynaptic) and the motor giant axon (postsynaptic) of the crayfish abdominal nerve cord have been investigated. Postsynaptic potential change in response to presynaptic volleys contains two components: a spike potential and a synaptic potential of very long time course. Amplitude of the synaptic potential is graded according to the number of active presynaptic axons. Conductance increase in the synaptic membrane endures over most of the period of potential change, and it is this rather than the "electrical time constant" of the membrane that in large measure determines the form of the synaptic potential. Temporal summation of synaptic potential occurs during repetitive presynaptic stimulation, and after such stimulation the rate of decay of synaptic potential is greatly slowed.     

Statistical Analysis of Membrane Potential Fluctuations: Relation with Presynaptic Spike Train

In a study of integration at the single neuron level, the relationships between the postsynaptic membrane potential and the presynaptic spike train were analyzed. Fluctuations in membrane potential of neurons in the visceral ganglion of Aplysia were measured and described by histograms. The histogram estimates the probability density function of the membrane potential. Comparisons were made among histograms when there was no synaptic input, and when there was a single input in which variations were made in the PSP (postsynaptic potential) sign, i.e. excitatory or inhibitory, and arrival statistics, e.g. slow or fast, regular, Poisson-like, or patterned. This was examined in cells where the membrane potential was constant and in cells in which there was spontaneous pacemaker activity. The form of the histogram depended on whether the neuron was spontaneously quiescent or a pacemaker, or whether it received presynaptic input and, if it did, on the sign and temporal characteristics of such input. From such histograms the mean firing rate of output spike trains can be predicted; additional information of a temporal nature is required, however, to predict features of the interval structure of the output train. Suggestions are made concerning the way the nervous system might utilize the information summarized in the membrane potential histogram.     

The Effect of Desflurane on Neuronal Communication at a Central Synapse

Although general anesthetics are thought to modify critical neuronal functions, their impact on neuronal communication has been poorly examined. We have investigated the effect induced by desflurane, a clinically used general anesthetic, on information transfer at the synapse between mossy fibers and granule cells of cerebellum, where this analysis can be carried out extensively. Mutual information values were assessed by measuring the variability of postsynaptic output in relationship to the variability of a given set of presynaptic inputs. Desflurane synchronized granule cell firing and reduced mutual information in response to physiologically relevant mossy fibers patterns. The decrease in spike variability was due to an increased postsynaptic membrane excitability, which made granule cells more prone to elicit action potentials, and to a strengthened synaptic inhibition, which markedly hampered membrane depolarization. These concomitant actions on granule cells firing indicate that desflurane re-shapes the transfer of information between neurons by providing a less informative neurotransmission rather than completely silencing neuronal activity.     

Electrical properties of neurons recorded from the rat supraoptic nucleus in vitro

The electrical properties of neurons in the supraoptic nucleus (so.n.) have been studied in the hypothalamic slice preparation by intracellular and extracellular recording techniques, with Lucifer Yellow CH dye injection to mark the recording site as being the so.n. Intracellular recordings from so.n. neurons revealed them to have an average membrane potential of -67 +/- 0.8 mV (mean +/- s.e.m.), membrane resistance of 145 +/- 9 M omega with linear current-voltage relations from 40 mV in the hyperpolarizing direction to the level of spike threshold in the depolarizing direction. Average cell time constant was 14 +/- 2.2 ms. So.n. action potentials ranged in amplitude from 55 to 95 mV, with a mean of 76 +/- 2 mV, and a spike width of 2.6 +/- 0.5 ms at 30% of maximal spike height. Both single spikes and trains of spikes were followed by a strong, long-lasting hyperpolarization with a decay fitted by a single exponential having a time constant of 8.6 +/- 1.8 ms. Action potentials could be blocked by 10(-6) M tetrodotoxin. Spontaneously active so.n. neurons were characterized by synaptic input in the form of excitatory and inhibitory postsynaptic potentials, the latter being apparently blocked when 4 M KCl electrodes were used. Both forms of synaptic activity were blocked by application of divalent cations such as Mg2+, Mn2+ or Co2+. 74% of so.n. neurons fired spontaneously at rates exceeding 0.1 spikes per second, with a mean for all cells of 2.9 +/- 0.2 s-1. Of these cells, 21% fired slowly and continuously at 0.1 - 1.0 s-1, 45% fired continuously at greater than 1 Hz, and the remaining 34% fired phasically in bursts of activity followed by silence or low frequency firing. Spontaneously firing phasic cells showed a mean burst length of 16.7 +/- 4.5 s and a silent period of 28.2 +/- 4.2 s. Intracellular recordings revealed the presence of slow variations in membrane potential which modified the neuron's proximity to spike threshold, and controlled phasic firing. Variations in synaptic input were not observed to influence firing in phasic cells.     

Motoneuron membrane potentials follow a time inhomogeneous jump diffusion process

Stochastic leaky integrate-and-fire models are popular due to their simplicity and statistical tractability. They have been widely applied to gain understanding of the underlying mechanisms for spike timing in neurons, and have served as building blocks for more elaborate models. Especially the Ornstein–Uhlenbeck process is popular to describe the stochastic fluctuations in the membrane potential of a neuron, but also other models like the square-root model or models with a non-linear drift are sometimes applied. Data that can be described by such models have to be stationary and thus, the simple models can only be applied over short time windows. However, experimental data show varying time constants, state dependent noise, a graded firing threshold and time-inhomogeneous input. In the present study we build a jump diffusion model that incorporates these features, and introduce a firing mechanism with a state dependent intensity. In addition, we suggest statistical methods to estimate all unknown quantities and apply these to analyze turtle motoneuron membrane potentials. Finally, simulated and real data are compared and discussed. We find that a square-root diffusion describes the data much better than an Ornstein–Uhlenbeck process with constant diffusion coefficient. Further, the membrane time constant decreases with increasing depolarization, as expected from the increase in synaptic conductance. The network activity, which the neuron is exposed to, can be reasonably estimated to be a threshold version of the nerve output from the network. Moreover, the spiking characteristics are well described by a Poisson spike train with an intensity depending exponentially on the membrane potential.     

Axonal speeding: shaping synaptic potentials in small neurons by the axonal membrane compartment

The role of the axonal membrane compartment in synaptic integration is usually neglected. We show here that in interneurons of the cerebellar molecular layer, where dendrites are so short that the somatodendritic domain can be considered isopotential, the axonal membrane contributes a significant part of the cell input capacitance. We examine the impact of axonal membrane on synaptic integration by cutting the axon with two-photon illumination. We find that the axonal compartment acts as a sink for signals generated at fast conductance synapses, thus increasing the initial decay rate of corresponding synaptic potentials over the value predicted from the resistance-capacitance (RC) product of the cell membrane; signals generated at slower synapses are much less affected. This mechanism sharpens the spike firing precision of fast glutamatergic inputs without resorting to multisynaptic pathways.     

Computational Consequences of Temporally Asymmetric Learning Rules: I. Differential Hebbian Learning

Temporally asymetric learning rules governing plastic changes in synaptic efficacy have recently been identified in physiological studies. In these rules, the exact timing of pre- and postsynaptic spikes is critical to the induced change of synaptic efficacy. The temporal learning rules treated in this article are approximately antisymmetric; the synaptic efficacy is enhanced if the postsynaptic spike follows the presynaptic spike by a few milliseconds, but the efficacy is depressed if the postsynaptic spike precedes the presynaptic spike. The learning dynamics of this rule are studied using a stochastic model neuron receiving a set of serially delayed inputs. The average change of synaptic efficacy due to the temporally antisymmetric learning rule is shown to yield differential Hebbian learning. These results are demonstrated with both mathematical analyses and computer simulations, and connections with theories of classical conditioning are discussed.     

Maturation of GABAergic Inhibition Promotes Strengthening of Temporally Coherent Inputs among Convergent Pathways

Spike-timing-dependent plasticity (STDP), a form of Hebbian plasticity, is inherently stabilizing. Whether and how GABAergic inhibition influences STDP is not well understood. Using a model neuron driven by converging inputs modifiable by STDP, we determined that a sufficient level of inhibition was critical to ensure that temporal coherence (correlation among presynaptic spike times) of synaptic inputs, rather than initial strength or number of inputs within a pathway, controlled postsynaptic spike timing. Inhibition exerted this effect by preferentially reducing synaptic efficacy, the ability of inputs to evoke postsynaptic action potentials, of the less coherent inputs. In visual cortical slices, inhibition potently reduced synaptic efficacy at ages during but not before the critical period of ocular dominance (OD) plasticity. Whole-cell recordings revealed that the amplitude of unitary IPSCs from parvalbumin positive (Pv+) interneurons to pyramidal neurons increased during the critical period, while the synaptic decay time-constant decreased. In addition, intrinsic properties of Pv+ interneurons matured, resulting in an increase in instantaneous firing rate. Our results suggest that maturation of inhibition in visual cortex ensures that the temporally coherent inputs (e.g. those from the open eye during monocular deprivation) control postsynaptic spike times of binocular neurons, a prerequisite for Hebbian mechanisms to induce OD plasticity.