Faithful representation of stimuli with a population of integrate-and-fire neurons

We consider a formal model of stimulus encoding with a circuit consisting of a bank of filters and an ensemble of integrate-and-fire neurons. Such models arise in olfactory systems, vision, and hearing. We demonstrate that bandlimited stimuli can be faithfully represented with spike trains generated by the ensemble of neurons. We provide a stimulus reconstruction scheme based on the spike times of the ensemble of neurons and derive conditions for perfect recovery. The key result calls for the spike density of the neural population to be above the Nyquist rate. We also show that recovery is perfect if the number of neurons in the population is larger than a threshold value. Increasing the number of neurons to achieve a faithful representation of the sensory world is consistent with basic neurobiological thought. Finally we demonstrate that in general, the problem of faithful recovery of stimuli from the spike train of single neurons is ill posed. The stimulus can be recovered, however, from the information contained in the spike train of a population of neurons.     

Model-based decoding, information estimation, and change-point detection techniques for multineuron spike trains

One of the central problems in systems neuroscience is to understand how neural spike trains convey sensory information. Decoding methods, which provide an explicit means for reading out the information contained in neural spike responses, offer a powerful set of tools for studying the neural coding problem. Here we develop several decoding methods based on point-process neural encoding models, or forward models that predict spike responses to stimuli. These models have concave log-likelihood functions, which allow efficient maximum-likelihood model fitting and stimulus decoding. We present several applications of the encoding model framework to the problem of decoding stimulus information from population spike responses: (1) a tractable algorithm for computing the maximum a posteriori (MAP) estimate of the stimulus, the most probable stimulus to have generated an observed single- or multiple-neuron spike train response, given some prior distribution over the stimulus; (2) a gaussian approximation to the posterior stimulus distribution that can be used to quantify the fidelity with which various stimulus features are encoded; (3) an efficient method for estimating the mutual information between the stimulus and the spike trains emitted by a neural population; and (4) a framework for the detection of change-point times (the time at which the stimulus undergoes a change in mean or variance) by marginalizing over the posterior stimulus distribution. We provide several examples illustrating the performance of these estimators with simulated and real neural data.     

Distortion of neural signals by spike coding

Analog neural signals must be converted into spike trains for transmission over electrically leaky axons. This spike encoding and subsequent decoding leads to distortion. We quantify this distortion by deriving approximate expressions for the mean square error between the inputs and outputs of a spiking link. We use integrate-and-fire and Poisson encoders to convert naturalistic stimuli into spike trains and spike count and inter-spike interval decoders to generate reconstructions of the stimulus. The distortion expressions enable us to compare these spike coding schemes over a large parameter space. We verify that the integrate-and-fire encoder is more effective than the Poisson encoder. The disparity between the two encoders diminishes as the stimulus coefficient of variation (CV) increases, at which point, the variability attributed to the stimulus overwhelms the variability attributed to Poisson statistics. When the stimulus CV is small, the interspike interval decoder is superior, as the distortion resulting from spike count decoding is dominated by a term that is attributed to the discrete nature of the spike count. In this regime, additive noise has a greater impact on the interspike interval decoder than the spike count decoder. When the stimulus CV is large, the average signal excursion is much larger than the quantization step size, and spike count decoding is superior.     

Derivation of the visual contrast response function by maximizing information rate

A graph of neural output as a function of the logarithm of stimulus intensity often produces an S-shaped function, which is frequently modeled by the hyperbolic ratio equation. The response of neurons in early vision to stimuli of varying contrast is an important example of this. Here, the hyperbolic ratio equation with a response exponent of two is derived exactly by considering the balance between information rate and the neural costs of making that information available, where neural costs are a function of synaptic strength and spike rate. The maximal response and semisaturation constant of the neuron can be related to the stimulus ensemble and therefore shift accordingly to exhibit contrast gain control and normalization.     

Efficient Markov chain Monte Carlo methods for decoding neural spike trains

Stimulus reconstruction or decoding methods provide an important tool for understanding how sensory and motor information is represented in neural activity. We discuss Bayesian decoding methods based on an encoding generalized linear model (GLM) that accurately describes how stimuli are transformed into the spike trains of a group of neurons. The form of the GLM likelihood ensures that the posterior distribution over the stimuli that caused an observed set of spike trains is log concave so long as the prior is. This allows the maximum a posteriori (MAP) stimulus estimate to be obtained using efficient optimization algorithms. Unfortunately, the MAP estimate can have a relatively large average error when the posterior is highly nongaussian. Here we compare several Markov chain Monte Carlo (MCMC) algorithms that allow for the calculation of general Bayesian estimators involving posterior expectations (conditional on model parameters). An efficient version of the hybrid Monte Carlo (HMC) algorithm was significantly superior to other MCMC methods for gaussian priors. When the prior distribution has sharp edges and corners, on the other hand, the "hit-and-run" algorithm performed better than other MCMC methods. Using these algorithms, we show that for this latter class of priors, the posterior mean estimate can have a considerably lower average error than MAP, whereas for gaussian priors, the two estimators have roughly equal efficiency. We also address the application of MCMC methods for extracting nonmarginal properties of the posterior distribution. For example, by using MCMC to calculate the mutual information between the stimulus and response, we verify the validity of a computationally efficient Laplace approximation to this quantity for gaussian priors in a wide range of model parameters; this makes direct model-based computation of the mutual information tractable even in the case of large observed neural populations, where methods based on binning the spike train fail. Finally, we consider the effect of uncertainty in the GLM parameters on the posterior estimators.     

The effects of spontaneous activity, background noise, and the stimulus ensemble on information transfer in neurons

Information theory is playing an increasingly important role in the analysis of neural data as it can precisely quantify the reliability of stimulus-response functions. Estimating the mutual information between a neural spike train and a time varying stimulus is, however, not trivial in practice and requires assumptions about the specific computations being performed by the neuron under study. Consequently, estimates of the mutual information depend on these assumptions and their validity must be ascertained in the particular physiological context in which experiments are carried out. Here we compare results obtained using different information measures that make different assumptions about the neural code (i.e. the way information is being encoded and decoded) and the stimulus ensemble (i.e. the set of stimuli that the animal can encounter in nature). Our comparisons are carried out in the context of spontaneously active neurons. However, some of our results are also applicable to neurons that are not spontaneously active. We first show conditions under which a single stimulus provides a good sample of the entire stimulus ensemble. Furthermore, we use a recently introduced information measure that is based on the spontaneous activity of the neuron rather than on the stimulus ensemble. This measure is compared to the Shannon information and it is shown that the two differ only by a constant. This constant is shown to represent the information that the neuron's spontaneous activity transmits about the fact that no stimulus is present in the animal's environment. As a consequence, the mutual information measure based on spontaneous activity is easily applied to stimuli that mimic those seen in nature, as it does not require a priori knowledge of the stimulus ensemble. Finally, we consider the effect of noise in the animal's environment on information transmission about sensory stimuli. Our results show that, as expected, such 'background' noise will increase the trial-to-trial variability of the neural response to repeated presentations of a stimulus. However, the same background noise can also increase the variability of the spike train and hence can lead to increased information transfer in the presence of background noise. Our study emphasizes how different assumptions can lead to different predictions for the information transmission of a neuron. Assumptions about the computations being performed by the system under study as well as the stimulus ensemble and background noise should therefore be carefully considered when applying information theory.     

Fast population coding

Uncertainty coming from the noise in its neurons and the ill-posed nature of many tasks plagues neural computations. Maybe surprisingly, many studies show that the brain manipulates these forms of uncertainty in a probabilistically consistent and normative manner, and there is now a rich theoretical literature on the capabilities of populations of neurons to implement computations in the face of uncertainty. However, one major facet of uncertainty has received comparatively little attention: time. In a dynamic, rapidly changing world, data are only temporarily relevant. Here, we analyze the computational consequences of encoding stimulus trajectories in populations of neurons. For the most obvious, simple, instantaneous encoder, the correlations induced by natural, smooth stimuli engender a decoder that requires access to information that is nonlocal both in time and across neurons. This formally amounts to a ruinous representation. We show that there is an alternative encoder that is computationally and representationally powerful in which each spike contributes independent information; it is independently decodable, in other words. We suggest this as an appropriate foundation for understanding time-varying population codes. Furthermore, we show how adaptation to temporal stimulus statistics emerges directly from the demands of simple decoding.     

Continuous speech recognition with sparse coding

Sparse coding is an efficient way of coding information. In a sparse code most of the code elements are zero; very few are active. Sparse codes are intended to correspond to the spike trains with which biological neurons communicate. In this article, we show how sparse codes can be used to do continuous speech recognition. We use the TIDIGITS dataset to illustrate the process. First a waveform is transformed into a spectrogram, and a sparse code for the spectrogram is found by means of a linear generative model. The spike train is classified by making use of a spike train model and dynamic programming. It is computationally expensive to find a sparse code. We use an iterative subset selection algorithm with quadratic programming for this process. This algorithm finds a sparse code in reasonable time if the input is limited to a fairly coarse spectral resolution. At this resolution, our system achieves a word error rate of 19%, whereas a system based on Hidden Markov Models achieves a word error rate of 15% at the same resolution.     

Multidimensional encoding strategy of spiking neurons

Neural responses in sensory systems are typically triggered by a multitude of stimulus features. Using information theory, we study the encoding accuracy of a population of stochastically spiking neurons characterized by different tuning widths for the different features. The optimal encoding strategy for representing one feature most accurately consists of narrow tuning in the dimension to be encoded, to increase the single-neuron Fisher information, and broad tuning in all other dimensions, to increase the number of active neurons. Extremely narrow tuning without sufficient receptive field overlap will severely worsen the coding. This implies the existence of an optimal tuning width for the feature to be encoded. Empirically, only a subset of all stimulus features will normally be accessible. In this case, relative encoding errors can be calculated that yield a criterion for the function of a neural population based on the measured tuning curves.     

Optimal tuning widths in population coding of periodic variables

We study the relationship between the accuracy of a large neuronal population in encoding periodic sensory stimuli and the width of the tuning curves of individual neurons in the population. By using general simple models of population activity, we show that when considering one or two periodic stimulus features, a narrow tuning width provides better population encoding accuracy. When encoding more than two periodic stimulus features, the information conveyed by the population is instead maximal for finite values of the tuning width. These optimal values are only weakly dependent on model parameters and are similar to the width of tuning to orientation or motion direction of real visual cortical neurons. A very large tuning width leads to poor encoding accuracy, whatever the number of stimulus features encoded. Thus, optimal coding of periodic stimuli is different from that of nonperiodic stimuli, which, as shown in previous studies, would require infinitely large tuning widths when coding more than two stimulus features.     

Inferring the capacity of the vector Poisson channel with a Bernoulli model

The capacity defines the ultimate fidelity limits of information transmission by any system. We derive the capacity of parallel Poisson process channels to judge the relative effectiveness of neural population structures. Because the Poisson process is equivalent to a Bernoulli process having small event probabilities, we infer the capacity of multi-channel Poisson models from their Bernoulli surrogates. For neural populations wherein each neuron has individual innervation, inter-neuron dependencies increase capacity, the opposite behavior of populations that share a single input. We use Shannon's rate-distortion theory to show that for Gaussian stimuli, the mean-squared error of the decoded stimulus decreases exponentially in both the population size and the maximal discharge rate. Detailed analysis shows that population coding is essential for accurate stimulus reconstruction. By modeling multi-neuron recordings as a sum of a neural population, we show that the resulting capacity is much less than the population's, reducing it to a level that can be less than provided with two separated neural responses. This result suggests that attempting neural control without spike sorting greatly reduces the achievable fidelity. In contrast, single-electrode neural stimulation does not incur any capacity deficit in comparison to stimulating individual neurons.     

Statistical significance of coincident spikes: count-based versus rate-based statistics.

Inspired by different conceptualizations of temporal neural coding schemes, there has been recent interest in the search for signs of precisely synchronized neural activity in the cortex. One method developed for this task is unitary-event analysis. This method tests multiple single neuron recordings for short epochs with significantly more coincident spikes than expected from independent neurons. We reformulated the statistical test underlying this method using a coincidence count distribution based on empirical spike counts rather than on estimated spike probabilities. In the case of two neurons, the requirement of stationary firing rates, originally imposed on both neurons, can be relaxed; only the rate of one neuron needs to be stationary, while the other may follow an arbitrary time course. By analytical calculations of the test power curves of the original and the revised method, we demonstrate that the test power can be increased by a factor of two or more in physiologically realistic regimes. In addition, we analyze the effective significance levels of both methods for neural firing rates ranging between 0.2 Hz and 30 Hz.     

Effect of lateral connections on the accuracy of the population code for a network of spiking neurons.

We study how neuronal connections in a population of spiking neurons affect the accuracy of stimulus estimation. Neurons in our model code for a one-dimensional orientation variable phi. Connectivity between two neurons depends on the absolute difference absolute value(phi - phi') between the preferred orientation of the two neurons. We derive an analytical expression of the activity profile for a population of neurons described by the spike response model with noisy threshold. We estimate the stimulus orientation and the trial-to-trial fluctuations using the population vector method. For stationary stimuli, uniform inhibitory connections produce a more reliable estimation of the stimulus than short-range excitatory connections with long-range inhibitions, although the latter interaction type produces a sharper tuning curve. These results are consistent with previous analytical studies of the Fisher information.     

Stimulus competition by inhibitory interference

When two stimuli are present in the receptive field of a V4 neuron, the firing rate response is between the weakest and strongest response elicited by each of the stimuli when presented alone (Reynolds, Chelazzi, & Desimone, 1999). When attention is directed toward the stimulus eliciting the strongest response (the preferred stimulus), the response to the pair is increased, whereas the response decreases when attention is directed to the other stimulus (the poor stimulus). When attention is directed to either of the two stimuli presented alone, the firing rate remains the same or increases slightly, but the coherence between the neuron's spike train and the local field potential can increase (Fries, Reynolds, Rorie, & Desimone, 2001). These experimental results were reproduced in a model of a V4 neuron under the assumption that attention modulates the activity of local interneuron networks. The V4 model neuron received stimulus-specific excitation from V2 and synchronous inhibitory inputs from two local interneuron networks in V4. Each interneuron network was driven by stimulus-specific excitatory inputs from V2 and was modulated by the activity of the frontal eye fields. Stimulus competition was present because of a delay in arrival time of synchronous volleys from each interneuron network. For small delays, the firing rate was close to the rate elicited by the preferred stimulus alone, whereas for larger delays, it approached the firing rate of the poor stimulus. When either stimulus was presented alone, the neuron's response was not altered by the change in delay, but could change due to modulation of the degree of synchrony of the corresponding interneuron network. The model suggests that top-down attention biases the competition between V2 columns for control of V4 neurons primarily by changing the relative timing of inhibition, whereas changes in the degree of synchrony of interneuron networks modulate the response to a single stimulus. The new mechanism proposed here for attentional modulation of firing rate, gain modulation by inhibitory interference, is likely to have more general applicability to cortical information processing.     

Spatiotemporal spike encoding of a continuous external signal

Interspike intervals of spikes emitted from an integrator neuron model of sensory neurons can encode input information represented as a continuous signal from a deterministic system. If a real brain uses spike timing as a means of information processing, other neurons receiving spatiotemporal spikes from such sensory neurons must also be capable of treating information included in deterministic interspike intervals. In this article, we examine functions of neurons modeling cortical neurons receiving spatiotemporal spikes from many sensory neurons. We show that such neuron models can encode stimulus information passed from the sensory model neurons in the form of interspike intervals. Each sensory neuron connected to the cortical neuron contributes equally to the information collection by the cortical neuron. Although the incident spike train to the cortical neuron is a superimposition of spike trains from many sensory neurons, it need not be decomposed into spike trains according to the input neurons. These results are also preserved for generalizations of sensory neurons such as a small amount of leak, noise, inhomogeneity in firing rates, or biases introduced in the phase distributions.     

Estimating spatio-temporal receptive fields of auditory and visual neurons from their responses to natural stimuli

We present a generalized reverse correlation technique that can be used to estimate the spatio-temporal receptive fields (STRFs) of sensory neurons from their responses to arbitrary stimuli such as auditory vocalizations or natural visual scenes. The general solution for STRF estimation requires normalization of the stimulus-response cross-correlation by the stimulus autocorrelation matrix. When the second-order stimulus statistics are stationary, normalization involves only the diagonal elements of the Fourier-transformed auto-correlation matrix (the power spectrum). In the non-stationary case normalization requires the entire auto-correlation matrix. We present modelling studies that demonstrate the feasibility and accuracy of this method as well as neurophysiological data comparing STRFs estimated using natural versus synthetic stimulus ensembles. For both auditory and visual neurons, STRFs obtained with these different stimuli are similar, but exhibit systematic differences that may be functionally significant. This method should be useful for determining what aspects of natural signals are represented by sensory neurons and may reveal novel response properties of these neurons.     

Quantifying neurotransmission reliability through metrics-based information analysis

We set forth an information-theoretical measure to quantify neurotransmission reliability while taking into full account the metrical properties of the spike train space. This parametric information analysis relies on similarity measures induced by the metrical relations between neural responses as spikes flow in. Thus, in order to assess the entropy, the conditional entropy, and the overall information transfer, this method does not require any a priori decoding algorithm to partition the space into equivalence classes. It therefore allows the optimal parameters of a class of distances to be determined with respect to information transmission. To validate the proposed information-theoretical approach, we study precise temporal decoding of human somatosensory signals recorded using microneurography experiments. For this analysis, we employ a similarity measure based on the Victor-Purpura spike train metrics. We show that with appropriate parameters of this distance, the relative spike times of the mechanoreceptors' responses convey enough information to perform optimal discrimination--defined as maximum metrical information and zero conditional entropy--of 81 distinct stimuli within 40 ms of the first afferent spike. The proposed information-theoretical measure proves to be a suitable generalization of Shannon mutual information in order to consider the metrics of temporal codes explicitly. It allows neurotransmission reliability to be assessed in the presence of large spike train spaces (e.g., neural population codes) with high temporal precision.     

Modelling the firing pattern of bullfrog vestibular neurons responding to naturalistic stimuli

We have developed a neural system identification method for fitting models to stimulus-response data, where the response is a spike train. The method involves using a general nonlinear optimisation procedure to fit models in the time domain. We have applied the method to model bullfrog semicircular canal afferent neuron responses during naturalistic, broad-band head rotations. These neurons respond in diverse ways, but a simple four parameter class of models elegantly accounts for the various types of responses observed.     

Estimating information rates with confidence intervals in neural spike trains

Information theory provides a natural set of statistics to quantify the amount of knowledge a neuron conveys about a stimulus. A related work (Kennel, Shlens, Abarbanel, & Chichilnisky, 2005) demonstrated how to reliably estimate, with a Bayesian confidence interval, the entropy rate from a discrete, observed time series. We extend this method to measure the rate of novel information that a neural spike train encodes about a stimulus--the average and specific mutual information rates. Our estimator makes few assumptions about the underlying neural dynamics, shows excellent performance in experimentally relevant regimes, and uniquely provides confidence intervals bounding the range of information rates compatible with the observed spike train. We validate this estimator with simulations of spike trains and highlight how stimulus parameters affect its convergence in bias and variance. Finally, we apply these ideas to a recording from a guinea pig retinal ganglion cell and compare results to a simple linear decoder.     

Quantifying variability in neural responses and its application for the validation of model predictions

A rate code assumes that a neuron's response is completely characterized by its time-varying mean firing rate. This assumption has successfully described neural responses in many systems. The noise in rate coding neurons can be quantified by the coherence function or the correlation coefficient between the neuron's deterministic time-varying mean rate and noise corrupted single spike trains. Because of the finite data size, the mean rate cannot be known exactly and must be approximated. We introduce novel unbiased estimators for the measures of coherence and correlation which are based on the extrapolation of the signal to noise ratio in the neural response to infinite data size. We then describe the application of these estimates to the validation of the class of stimulus-response models that assume that the mean firing rate captures all the information embedded in the neural response. We explain how these quantifiers can be used to separate response prediction errors that are due to inaccurate model assumptions from errors due to noise inherent in neuronal spike trains.