Sparse ensemble neural code for a complete vocal repertoire

The categorization of animal vocalizations into distinct behaviorally relevant groups for communication is an essential operation that must be performed by the auditory system. This auditory object recognition is a difficult task that requires selectivity to the group identifying acoustic features and invariance to renditions within each group. We find that small ensembles of auditory neurons in the forebrain of a social songbird can code the bird's entire vocal repertoire (∼10 call types). Ensemble neural discrimination is not, however, correlated with single unit selectivity, but instead with how well the joint single unit tunings to characteristic spectro-temporal modulations span the acoustic subspace optimized for the discrimination of call types. Thus, akin to face recognition in the visual system, call type recognition in the auditory system is based on a sparse code representing a small number of high-level features and not on highly selective grandmother neurons.

High-capacity auditory memory for vocal communication in a social songbird

Effective vocal communication often requires the listener to recognize the identity of a vocalizer, and this recognition is dependent on the listener's ability to form auditory memories. We tested the memory capacity of a social songbird, the zebra finch, for vocalizer identities using conditioning experiments and found that male and female zebra finches can remember a large number of vocalizers (mean, 42) based solely on the individual signatures found in their songs and distance calls. These memories were formed within a few trials, were generalized to previously unheard renditions, and were maintained for up to a month. A fast and high-capacity auditory memory for vocalizer identity has not been demonstrated previously in any nonhuman animals and is an important component of vocal communication in social species.

Neural response to bird?s own song and tutor song in the zebra finch field L and caudal mesopallium

Neurons in the song system nuclei of songbirds exhibit a strong preference for the sound of the bird's own song relative to that of conspecific songs. This selectivity is observed in the high vocal center and the nucleus interface of the nidopallium, two song nuclei that receive input from the bird's auditory system. To investigate the role of the auditory system in generating the selective responses observed in the song system, we recorded auditory responses in the zebra finch primary auditory forebrain, field L, and in a secondary auditory area, the caudal mesopallium. Field L and caudal mesopallium project directly or indirectly to the high vocal center and nucleus interface of the nidopallium and are presumed to provide substantial auditory input to the song system. We found that, on average, neurons in field L and caudal mesopallium did not show positive selective responses for the bird's own song or tutor song relative to conspecific song. Moreover, there were no particular sub-areas in the auditory telencephalon that were relatively more selective than the average. The selectivity for the bird's own song would therefore be restricted to song nuclei and would arise in one processing step, potentially found at the interface between the auditory and the song systems.

Feature analysis of natural sounds in the songbird auditory forebrain

Although understanding the processing of natural sounds is an important goal in auditory neuroscience, relatively little is known about the neural coding of these sounds. Recently we demonstrated that the spectral temporal receptive field (STRF), a description of the stimulus-response function of auditory neurons, could be derived from responses to arbitrary ensembles of complex sounds including vocalizations. In this study, we use this method to investigate the auditory processing of natural sounds in the birdsong system. We obtain neural responses from several regions of the songbird auditory forebrain to a large ensemble of bird songs and use these data to calculate the STRFs, which are the best linear model of the spectral-temporal features of sound to which auditory neurons respond. We find that these neurons respond to a wide variety of features in songs ranging from simple tonal components to more complex spectral-temporal structures such as frequency sweeps and multi-peaked frequency stacks. We quantify spectral and temporal characteristics of these features by extracting several parameters from the STRFs. Moreover, we assess the linearity versus nonlinearity of encoding by quantifying the quality of the predictions of the neural responses to songs obtained using the STRFs. Our results reveal successively complex functional stages of song analysis by neurons in the auditory forebrain. When we map the properties of auditory forebrain neurons, as characterized by the STRF parameters, onto conventional anatomical subdivisions of the auditory forebrain, we find that although some properties are shared across different subregions, the distribution of several parameters is suggestive of hierarchical processing.

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.

Extraction of sensory parameters from a neural map by primary sensory interneurons

We examine the anatomical basis for the representation of stimulus parameters within a neural map and examine the extraction of these parameters by sensory interneurons (INs) in the cricket cercal sensory system. The extraction of air current direction by these sensory interneurons can be understood largely in terms of the anatomy of the system. There are two critical anatomical constraints. (1) The arborizations of afferents with similar directional tuning properties are located near each other within the neural map. Therefore, a continuous variation in stimulus direction causes a continuous variation in the spatial pattern of activation. (2) The restriction of the synaptic connections of an interneuron to a unique set of afferents results from the unique anatomy of that interneuron: its dendritic arbors are located within restricted regions of the afferent map containing afferents with a limited subset of directional sensitivities. The functional organization of the set of four interneurons studied here is equivalent to a Cartesian coordinate system for computing the stimulus direction vector. For any air current stimulus direction, the firing rates of the active interneurons could be decoded as Cartesian coordinates by neurons at successive processing stages. The implications of this Cartesian coordinate system are discussed with respect to optimal coding strategies and developmental constraints on the cellular implementation of this coding scheme.

Spectral-temporal receptive fields of nonlinear auditory neurons obtained using natural sounds

The stimulus-response function of many visual and auditory neurons has been described by a spatial-temporal receptive field (STRF), a linear model that for mathematical reasons has until recently been estimated with the reverse correlation method, using simple stimulus ensembles such as white noise. Such stimuli, however, often do not effectively activate high-level sensory neurons, which may be optimized to analyze natural sounds and images. We show that it is possible to overcome the simple-stimulus limitation and then use this approach to calculate the STRFs of avian auditory forebrain neurons from an ensemble of birdsongs. We find that in many cases the STRFs derived using natural sounds are strikingly different from the STRFs that we obtained using an ensemble of random tone pips. When we compare these two models by assessing their predictions of neural response to the actual data, we find that the STRFs obtained from natural sounds are superior. Our results show that the STRF model is an incomplete description of response properties of nonlinear auditory neurons, but that linear receptive fields are still useful models for understanding higher level sensory processing, as long as the STRFs are estimated from the responses to relevant complex stimuli.

Information theory and neural coding

Information theory quantifies how much information a neural response carries about the stimulus. This can be compared to the information transferred in particular models of the stimulus-response function and to maximum possible information transfer. Such comparisons are crucial because they validate assumptions present in any neurophysiological analysis. Here we review information-theory basics before demonstrating its use in neural coding. We show how to use information theory to validate simple stimulus-response models of neural coding of dynamic stimuli. Because these models require specification of spike timing precision, they can reveal which time scales contain information in neural coding. This approach shows that dynamic stimuli can be encoded efficiently by single neurons and that each spike contributes to information transmission. We argue, however, that the data obtained so far do not suggest a temporal code, in which the placement of spikes relative to each other yields additional information.

Temporal and spectral sensitivity of complex auditory neurons in the nucleus HVc of male zebra finches

Complex vocalizations, such as human speech and birdsong, are characterized by their elaborate spectral and temporal structure. Because auditory neurons of the zebra finch forebrain nucleus HVc respond extremely selectively to a particular complex sound, the bird's own song (BOS), we analyzed the spectral and temporal requirements of these neurons by measuring their responses to systematically degraded versions of the BOS. These synthetic songs were based exclusively on the set of amplitude envelopes obtained from a decomposition of the original sound into frequency bands and preserved the acoustical structure present in the original song with varying degrees of spectral versus temporal resolution, which depended on the width of the frequency bands. Although both excessive temporal or spectral degradation eliminated responses, HVc neurons responded well to degraded synthetic songs with time-frequency resolutions of approximately 5 msec or 200 Hz. By comparing this neuronal time-frequency tuning with the time-frequency scales that best represented the acoustical structure in zebra finch song, we concluded that HVc neurons are more sensitive to temporal than to spectral cues. Furthermore, neuronal responses to synthetic songs were indistinguishable from those to the original BOS only when the amplitude envelopes of these songs were represented with 98% accuracy. That level of precision was equivalent to preserving the relative time-varying phase across frequency bands with resolutions finer than 2 msec. Spectral and temporal information are well known to be extracted by the peripheral auditory system, but this study demonstrates how precisely these cues must be preserved for the full response of high-level auditory neurons sensitive to learned vocalizations.

Functional organization of a neural map in the cricket cercal sensory system

Directionally selective mechanosensory afferents in the cricket cercal sensory system form a map of air current direction in the terminal abdominal ganglion. The global organization of this map was revealed by studying the anatomical relationships between an ensemble of sensory afferents that represented the entire range of receptor hair directional sensitivities on the sensory epithelium. The shapes and three-dimensional positions of the terminal arborizations of these cells were highly conserved across animals. Afferents with similar directional sensitivities arborized near each other within the map, and their terminal arborizations showed significant anatomical overlap. There was a clear global organization pattern of afferents within the map: they were organized into a spiral shape, with stimulus direction mapped continuously around the spiral. These results demonstrate that this map is not formed via a direct point-to-point topographic projection from the sensory epithelium to the CNS. Rather, the continuous representation of air current direction is synthesized within the CNS via an anatomical reorganization of the afferent terminal arbors. The arbors are reorganized according to a functional property that is independent of the location of the mechanoreceptor in the epithelium. The ensemble data were used to derive predictions of the patterns of steady-state excitation throughout the map for different directional stimuli. These images represent quantitative and testable predictions of functional characteristics of the entire neural map.

Representation of sensory information in the cricket cercal sensory system. I. Response properties of the primary interneurons

1. Six different types of primary wind-sensitive interneurons in the cricket cercal sensory system were tested for their sensitivity to the orientation and peak velocity of unidirectional airflow stimuli. 2. The cells could be grouped into two distinct classes on the basis of their thresholds and static sensitivities to airflow velocity. 3. Four interneurons (the right and left 10-2 cells and the right and left 10-3 cells) made up one of the two distinct velocity sensitivity classes. The mean firing frequencies of these interneurons were proportional to the logarithm of peak stimulus velocity over the range from 0.02 to 2.0 cm/s. 4. The other two interneurons studied (left and right 9-3) had a higher air-current velocity threshold, near the saturation level of the 10-2 and 10-3 interneurons. The slope of the velocity sensitivity curve for the 9-3 interneurons was slightly greater than that for the 10-2 and 10-3 interneurons, extending the sensitivity range of the system as a whole to at least 100 cm/s. 5. All of the interneurons had broad, symmetrical, single-lobed directional sensitivity tuning curves that could be accurately represented as truncated sine waves with 360 degree period. 6. The four low-threshold interneurons (i.e., left and right 10-2 and 10-3) had peak directional sensitivities that were evenly spaced around the horizontal plane, and their overlapping tuning curves covered all possible air-current stimulus orientations. The variance in the cells' responses to identical repeated stimuli varied between approximately 10% at the optimal stimulus orientations and approximately 30% at the zero-crossing orientations. 7. The two higher threshold interneurons (left and right 9-3) had broader directional sensitivity curves and wider spacing, resulting in reduced overlap with respect to the low-threshold class.

Representation of sensory information in the cricket cercal sensory system. II. Information theoretic calculation of system accuracy and optimal tuning-curve widths of four primary interneurons

1. Principles of information theory were used to calculate the limit of accuracy achievable by a subset of the wind-sensitive primary interneurons in the cricket cercal sensory system. For these calculations, an ensemble of four neurons was treated as an information channel, which encoded the direction of air-current stimuli for a defined range of air-current velocities. The specific information theoretic parameter that was calculated was the "transin-formation" or "mutual information" between the air-current directions and the neuronal spike trains, which were characterized in the preceding report. Under the assumptions used for these calculations, the ensemble of four interneurons was demonstrated to be capable of encoding between 4.2 and 3.5 bits of information about wind direction. This corresponds to an average directional accuracy of 4.7 and 7.7 degrees, respectively. 2. The same principles were applied to estimate the extent to which any variation in the width of the tuning curves would affect the transfer of information. As the widths of simulated tuning curves were varied, the mean ensemble accuracy showed a clear global maximum. This maximum corresponds to tuning curves widths of 110 degrees wide (at half maximum), which was remarkably close to the actual mean widths of the tuning curves observed in the cricket of 130 degrees. 3. The effect of varying the parametric "spacing" of the tuning curves within the stimulus range was also examined through a series of simulations. The configuration allowing the maximum information transfer corresponded to equal spacing of the tuning curves around the stimulus range (i.e., 90 degrees separation of peak sensitivity points). This theoretically optimum spacing corresponded exactly to the values observed in the experiments presented in the preceding report. 4. These simulations also showed that the degradation in the accuracy resulting from a shift in the tuning-curve spacing would depend on the plasticity of the higher order decoder of directional information. If there were no plasticity in the interneurons making up the higher order decoder, then the accuracy would be degraded by 50% for a mean tuning-curve shift of only 3.5 degrees. However, if the higher order decoding network were capable of being reoptimized to any arbitrary shift in tuning curves, the degradation in attainable accuracy would be much less severe as shifts of up to 10 degrees would result in virtually no degradation in the accuracy. 5. From these results, two general conclusions can be drawn about the coding of specific stimulus parameters by arrays of sensory cells.(ABSTRACT TRUNCATED AT 400 WORDS)