Running Head Cortical Receptive Fields and Discrimination Corresponding Author

Journal of Neurophysiology

Report

Delayed Inhibition in Cortical Receptive Fields and the Discrimination of Complex Stimuli Rajiv Narayan, Ayla Ergün and Kamal Sen

Hearing Research Center, Department of Biomedical Engineering

Center for Biodynamics & Program in Mathematical and Computational NeuroscienceBoston University, Boston, MA 02215

Running Head: Cortical Receptive Fields and Discrimination

Corresponding Author:Kamal Sen

Department of Biomedical EngineeringBoston University44 Cummington St.Boston, MA 02215Email: kamalsen@bu.eduPhone: 617-353-5919Fax: 617-353-6766

1

Copyright © 2005 by the American Physiological Society.

Abstract

Although, auditory cortex is thought to play an important role in processing complex natural sounds such as speech and animal vocalizations, the specific functional roles of cortical receptive fields (RFs) remain unclear. Here, we investigate the relationship between a behaviorally important function: the discrimination of natural sounds, and the structure of cortical RFs. We examine this problem in the model system of songbirds, using a computational approach. First, we construct model neurons based on the Spectral Temporal Receptive Field (STRF), a widely used description of auditory cortical RFs. We focus on delayed inhibitory STRFs, a class of STRFs experimentally observed in primary auditory cortex (ACx) and its analog in songbirds (field L), which consist of an excitatory subregion and a delayed inhibitory subregion co-tuned to a characteristic frequency. We then quantify the discrimination of birdsongs by model neurons, examining both the dynamics and temporal resolution of discrimination, using a recently proposed Spike Distance Metric (SDM). We find that single model neurons with delayed inhibitory STRFs, can discriminate accurately between songs. Discrimination improves dramatically when the temporal structure of the neural response at fine timescales is considered. When we compare discrimination by model neurons with and without the inhibitory subregion, we find that the presence of the inhibitory subregion can improve discrimination. Finally, we model a cortical microcircuit with delayed synaptic inhibition, a candidate mechanism underlying delayed inhibitory STRFs, and show that blocking inhibition in this model circuit degrades discrimination.

Keywords: auditory cortex, receptive field, discrimination, natural sounds, field L

2

Introduction

Auditory cortex is thought to play an important role in processing complex natural sounds

such as human speech and animal vocalizations (Fitch et al. 1997; Rauschecker 1998). Recent studies have described the structure of neuronal receptive fields (RFs) in primary auditory cortex (ACx), employing natural as well as complex naturalistic sounds (Depireux et al. 2001; Elhilali et al. 2004; Linden et al. 2003; Machens et al. 2004; Miller et al. 2001, 2002). Although such studies have revealed new aspects of cortical representations for complex sounds, the functional roles of cortical neurons in processing natural sounds remain unclear.

The combination of well understood vocal communication behaviors and identified neural

circuits that mediate the perception, production and learning of vocal communication sounds makes the songbird system attractive for studying the functional roles of neural substrates (Brainard and Doupe 2002; Konishi 1985). Here, we explore the role of neurons in field L, the ACx analogue in songbirds, in a behaviorally important function: discriminating between conspecific songs (songs of the bird’s own species).

Previous experiments suggest a link between discrimination of simple synthetic sounds e.g.,

tones, and RFs (Edeline and Weinberger 1993; Fritz et al. 2003; Recanzone et al. 1993). Yet, how RF structure may influence the discrimination of natural sounds remains poorly understood. We take a computational approach to relate song discrimination to RF structure in field L. First, we model auditory neurons in field L based on previous experimental measurements of Spectral Temporal Receptive Fields (STRFs), a quantitative description of the receptive fields of auditory neurons (Sen et al. 2001; Theunissen et al. 2000). Here, we focus on delayed inhibitory STRFs, which contain a single excitatory and a single inhibitory region localized around a characteristic frequency. We chose these STRFs for three reasons. First, these STRFs are simple in structure, providing a good starting point for computational analysis. Second, these are a general class of RFs, found in field L, ACx of several species (Depireux et al. 2001; Elhilali et al. 2004; Linden et al. 2003; Machens et al. 2004; Miller et al. 2001, 2002; Sen et al. 2001; Theunissen et al. 2000), as well as other sensory modalities such as vision and somatosensation (DeAngelis et al. 1995; Ghazanfar and Nicolelis 2001). Thus, the analysis of such RFs may provide general insights into cortical sensory processing of natural stimuli. Third, a candidate microcircuit underlying such RFs, involving delayed synaptic inhibition, has been identified by recent experiments (Cruikshank et al 2002; Tan et al. 2004; Wehr and Zador 2003). We take advantage of this to explicitly model this circuit, and compare the results with the STRF model. To characterize neural discrimination of songs, we use a recently proposed theoretical measure, the Spike Distance Metric (SDM), to quantify the dynamics and temporal resolution of discrimination (Machens et al. 2003; van Rossum 2001). Finally, we generate model STRFs (Qiu et al. 2003), to assess the resulting changes in discrimination.

We use this approach to address several questions. How accurately do model neurons

discriminate between natural sounds? How does the accuracy of discrimination evolve over time as the neural response accumulates? Can STRF structure be related to discrimination? Specifically, how does the delayed inhibitory region in the STRF affect the discrimination of songs? Finally, we model a cortical microcircuit, based on delayed synaptic inhibition, which may underlie such STRFs, and investigate the effects of blocking inhibition in this model circuit on the discrimination of natural sounds.

Methods

We used a recently described parametric description to model STRFs (Qiu et al. 2003). The

spectral and temporal profiles of the STRF were modeled using Gabor functions. The difference in our study is that we used the sum of two Gabor functions to model the temporal profile, which allowed us to alter the properties of the excitatory and inhibitory subregions of the STRF independently. The best modulation frequency (BMF) for the delayed inhibitory STRF we used was 40 Hz.Other BMFs in the range of 10-40 Hz were also tested with similar results. We also modeled narrow and broadband STRFs with identical temporal profiles but different spectral bandwidths (the half-maximal bandwidth ranged from 1-4 kHz).

We constructed a computational model of a cortical neuron, based on the STRF. As input to

the model neurons, we used 20 zebra finch songs, which have previously been used to experimentally

3

characterize the STRFs of neurons in field L (Theunissen et al. 2000; Sen et al. 2001). The distribution of song durations ranged from 0.8 to 4.4 s (mean=2.0 s, median=1.8 s, mode=1.4 s).An estimate of the firing rate of the neuron was obtained by convolving the song spectrogram with the STRF (Theunissen et al. 2000; Sen et al. 2001). The firing rate was then used to drive a stochastic spike train generator to obtain several “trials” of spike trains (Dayan and Abbott 2001). The stochastic model was an inhomogeneous Poisson process with an absolute and a relative refractory period, with values of 2 ms and 5 ms respectively (Berry and Meister 1998).

We employed the SDM to quantify the dissimilarity between pairs of spike trains (van

Rossum 2001). This metric is computed as the Euclidean distance between a pair of spike trains, which have been filtered using a decaying exponential kernel with time constant󰁗. The parameter 󰁗can be varied to examine discrimination over different timescales of the neural response. We used a classification scheme based on the SDM to quantify the neural discrimination of songs (Machens et al. 2003). 10 trials of spike trains were obtained from the model for each of the 20 songs aligned at their onsets. A template spike train was chosen for each of the songs, and remaining spike trains were assigned to the song with the closest template based on the spike distance measure. This procedure was repeated 100 times for different templates. The percentage of correctly classified songs (% correct) was used as a measure of discrimination. The chance level for classification was 5%, since a spike train could be assigned to 1 of 20 songs. To compare discrimination between models with and without an inhibitory subregion, we constrained the average firing rate, over all songs in the two models to be equal. The average rate was 17 Hz, which is within the range of experimentally observed values in field L (Sen et al. 2001). To verify that our results were metric independent, we also employed an alternate cost-based distance metric (Victor and Purpura 1997).

We constructed a model cortical microcircuit, similar to a model recently described by Wehr

and Zador, 2003. The model consisted of an excitatory (E) input neuron and an inhibitory (I) input neuron, connected to an excitatory (E) output neuron (Fig. 3a). Spike trains of the E input neuron were generated using a purely excitatory STRF and the stochastic spike generator described earlier. The spike trains of the I neuron were generated by delaying and jittering the excitatory spike train. The delay was 2.5 ms, and the jitter was a random number drawn from a Gaussian distribution with mean 0 and standard deviation 1 ms. The output E neuron was a standard integrate-and-fire neuron, with excitatory and inhibitory synaptic conductances. Synaptic inputs were modeled using alpha functions (Dayan and Abbott 2001), with time constants of 1.5 ms for the excitatory synapse, and 12 ms for the inhibitory synapse. The strengths of the excitatory and inhibitory synapses were controlled using two synaptic weights,we andwi, withwiset to 0 for the circuit without inhibition. The average firing rate in both models was constrained to be equal at 17 Hz, by scaling the firing rate of the input E neuron.

Results

A Single Model Neuron Discriminates Accurately between Songs

We quantified discrimination of songs by a model neuron with a delayed inhibitory STRF (Fig

1a, inset), using a classification method based on the SDM (see Methods). Figure 1a illustrates the accuracy of song discrimination (% correct) as a function of spike train duration for󰁗=10ms. Song discrimination increases rapidly between 0-500 ms of song duration, reaching an accuracy level of approximately 70%, after which it improves more gradually and reaches a maximal accuracy of 97%. We obtained similar results with an alternate cost-based distance metric (Victor and Purpura 1997).

Discrimination is Optimal at a Fine Temporal Resolution

To compare discrimination at different timescales of the neural response, we varied the time-constant in the SDM (see Methods). At small timescales the SDM acts like a “coincidence detector”, with small differences in spike timing contributing to the distance, whereas at long timescales the metric acts like a “rate difference counter”, where average firing rates contribute to the distance (van Rossum, 2001). Figure 1b shows the discrimination performance for the model neuron with the delayed inhibitory STRF at different timescales. At a very long timescale of 1000 ms, discrimination accuracy is poor (20%). As the timescale is decreased, the temporal structure of the neural response is considered at increasingly finer resolutions and discrimination improves significantly, reaching a

4

maximal value of 92% at 10 ms. The discrimination accuracy is above 80% for timescales ranging from 4 ms to 25 ms, and drops sharply for values outside this range.

The Delayed Inhibitory Subregion Improves Discrimination

Next, we generated STRFs without an inhibitory subregion, and compared the discrimination

curves for model neurons with and without an inhibitory subregion in the STRF (Fig. 2 insets). The discrimination curve for the STRF with the inhibitory region rises more rapidly and reaches a higher final level of accuracy compared to the STRF without the inhibitory region (Fig. 2). Both curves plateau at about 1300 ms, when the inhibitory STRF model achieves an accuracy level of 97% while the curve without inhibition results in an accuracy level of only 67%.

Blocking Delayed Synaptic Inhibition in a Model Circuit Degrades Discrimination

Finally, we investigated discrimination in a model circuit, which has been proposed to

underlie delayed inhibitory STRFs (Fig. 3a). This circuit consisted of an excitatory neuron receiving two synaptic inputs: a direct excitatory input and an indirect i.e., disynaptic inhibitory input. Figure 3b shows the discrimination curves for the output excitatory neuron, with and without inhibition (see Methods). Discrimination performance with inhibition plateaus at about 1400 ms achieving 97% accuracy, whereas performance without inhibition improves at a much slower rate reaching 72% accuracy at 3000 ms.

Discussion

A Candidate Cortical Microcircuit Underlying Delayed Inhibitory STRFs

In ACx, the timing of inhibitory conductance has been shown to be delayed relative to

excitatory conductance (Tan et al. 2004; Wehr and Zador 2003). Previous studies have suggested that such a scheme may provide a mechanistic substrate for delayed inhibitory cortical STRFs (Miller et al. 2001). We constructed a simple model of such a cortical microcircuit with delayed inhibition, and investigated the discrimination of natural sounds, with and without inhibition. We found that the presence of delayed inhibition in this circuit, improved the discrimination of songs in a manner similar to that observed for the model with the delayed inhibitory STRF.

Delayed inhibition in field L appears to be mediated by a microcircuit similar to that in ACx,

consisting of local GABAergic interneurons driven by feed-forward excitation from the thalamus (Muller and Scheich 1987, 1988). Recent experiments have revealed additional features of delayed inhibition in ACx. The delay between excitatory and inhibitory synaptic conductances has been measured directly using intracellular recordings (mean delay ~2-4 ms; Cruikshank et al. 2002; Tan et al. 2004; Wehr and Zador 2003). These studies suggest a GABAA receptor mediated mechanism underlying delayed inhibition (Cruikshank et al. 2002; Tan et al. 2004), although a longer-lasting suppression, which may be mediated by synaptic depression, has also been reported (Wehr and Zador, 2005). Similar intracellular measurements in field L will allow a more detailed comparison between the micro-circuitry of avian and mammalian systems.

The importance of ACx in processing natural sounds such as human speech suggests

specialized cortical mechanisms for efficiently processing such sounds. Intracortical inhibition may play an important role in the construction of specific spectral-temporal filters e.g., delayed inhibitory STRFs. Consistent with this idea, recent studies in ACx have suggested that an important difference between thalamic and cortical STRFs is the structure of the inhibitory region (Miller et al. 2001). Importance of Temporal Structure in Complex Sounds and Neural Responses

Clearly, the efficacy of particular classes of STRFs in processing natural sounds depends on

the spectral and temporal cues present in the sounds. Temporal cues play a particularly important role in speech recognition (Shannon et al. 1995). Delayed inhibitory STRFs may provide a neural substrate for extracting temporal cues e.g., onsets, from sounds with time-varying envelopes e.g., vocal communication sounds, music or environmental sounds. Similar cortical RFs in humans may allow better discrimination of sounds in speech and music, which differ in temporal attributes e.g., voice onset time, formant transition duration, prosody and rhythm.

5

Our study also indicates the potential importance of the fine temporal structure of neural

responses in cortical discrimination. We found that for the delayed inhibitory STRFs, best discrimination was obtained at a relatively fine temporal resolution (~10 ms). Previous studies have revealed precise timing in auditory cortical responses (DeWeese et al. 2003; Elhilali et al. 2004; Heil 1997; Lu et al. 2001; Wehr and Zador 2003). Our results show that information on a fine timescale is available for a behaviorally important function: discriminating the sounds of conspecifics. Several factors may shape the optimal temporal resolution for discrimination. These include extrinsic time-scales present in the natural sounds (e.g., the time-scales associated with the sound onsets and amplitude modulations (AM)) as well as the time-scales intrinsic to the neuron (e.g., time scales in the RF, time scales associated with refractoriness). An intrinsic time-scale that might influence the optimal temporal resolution is the integration time-scale, or the inverse of the BMF. Specifically, neurons with larger integration time-scales may have larger optimal temporal resolution. We did not find such a direct relationship between the BMF and the optimal temporal resolution (data not shown, see Methods for range of BMFs tested). Thus, discrimination resolution may depend primarily on extrinsic time-scales e.g., time-scales associated with onsets or other intrinsic time-scales. Manipulating the statistics of natural sounds e.g., onsets and AM, and examining the resulting changes in the optimal temporal resolution may clarify the relative contributions of extrinsic vs. intrinsic time-scales towards shaping the temporal resolution for discrimination.

Scope and Future Directions

Delayed inhibitory RFs are found in field L, ACx as well as other sensory cortices (DeAngelis

et al. 1995; Depireux et al. 2001; Elhilali et al. 2004; Ghazanfar and Nicolelis 2001; Linden et al. 2003; Machens et al. 2004; Miller et al. 2001, 2002; Sen et al. 2001; Theunissen et al. 2000). Thus, our results suggest that delayed inhibition in RFs may provide a general cortical mechanism for efficient discrimination of dynamic natural stimuli across multiple modalities. However, such RFs are one of many classes of cortical RFs observed experimentally.By itself, a narrowband onset detector, such as the delayed inhibitory STRFs considered here, is unlikely to be sufficient for all discrimination tasks. For example, a problem arises when there is insufficient power in the limited frequency range to which the narrowband onset detector is sensitive. A population of delayed inhibitory STRFs centered at different frequencies, or a STRF with broader spectral bandwidth i.e., broadband onset detectors, may be more robust. Broadband onset detectors are more common in field L relative to narrowband onset detectors (Theunissen et al. 2004). Broadband onset detectors produced discrimination very similar to narrowband onset detectors (data not shown, see Methods for parameters). Thus, both broadband and narrowband onset detectors may contribute towards accurate discrimination of birdsongs. Discrimination tasks involving more complex spectral cues may also require spectral edge detectors e.g., STRFs with an excitatory subregion and inhibitory sidebands, or complex STRFs with excitatory and inhibitory subregions at multiple frequencies. These different classes of STRFs have been experimentally observed in field L and ACx (Sen et al. 2001; Theunissen et al. 2004). In future, it will be important to explore the discrimination capabilities of other RF classes and RF populations and to relate neural discrimination to behavior.

6

Understanding the Cortical Code: Stimulus Statistics vs. Behavior

A theoretical proposal that has been influential in the visual system is that the goal of sensory

processing is to construct efficient representations for natural stimuli (Attneave 1954; Barlow 1961; Simoncelli and Olshausen 2001). Such a theory predicts a relationship between receptive field properties of visual neurons and the statistics of natural images. Experimental and theoretical studies in vision suggest that efficient representations for natural images can be found in the primary visualcortex (Olshausen and Field 1996; Vinje and Gallant 2000).

A similar line of inquiry has recently been initiated for the auditory system (Attias and

Schreiner 1998; Lewicki 2002; Singh and Theunissen 2003). Interestingly, experimental and theoretical studies, have found statistically efficient representations of natural sounds as early as the auditory nerve (Lewicki 2002; Rieke et al. 1995). This finding raises an intriguing question: if efficient sensory codes are established at the periphery, what are the computational roles of subsequent auditory processing? We propose that a goal of subsequent stages of processing may be to build upon efficient sensory representations to create new representations that are well suited for behaviorally important tasks e.g., discrimination and recognition of species-specific and other complex sounds.

Both intra-cortical and pre-cortical auditory processing may contribute to the generation of

efficient behavioral representations. In the avian auditory system, selectivity for conspecific songs relative to synthetic sounds gradually increases from the midbrain nucleus mesencephalicus lateralis, pars dorsalis (MLd), the avian analog of inferior colliculus, to field L, to the caudal mesopallium (CM), which is analogous to secondary auditory cortex (Grace et al. 2003; Hsu et al. 2004; Theunissen et al. 2004). This pattern suggests an intra-cortical contribution to increasing selectivity for conspecific sounds. Secondary cortical areas CM and the caudal nidopallium (NCM) respond to song familiarity, suggesting a role in individual vocal recognition (Gentner 2004; Gentner and Margoliash 2003; Mello et al. 2004; Mello et al. 1992). The process of recognition can be conceptualized in two stages: discrimination followed by association. Thus, improvements in song discrimination in field L, may assist in song recognition in secondary cortical areas. Some aspects of the cortical code may also be inherited from pre-cortical processing. For example, onset detector type STRFs have recently been reported in MLd (Woolley and Theunissen 2005), raising the possibility that similar STRFs found in field L may be inherited from MLd. Localizing where efficient representations for behavior originate will require further comparative analysis of response properties along hierarchical stages of auditory processing. Regardless of intra-cortical or pre-cortical origins, such representations are likely to transform the cortical code. If so, understanding the cortical code will require an integrative analysis of receptive fields, stimulus statistics as well as behavior.

Acknowledgments: We thank Allison Doupe and the Doupe Lab for support and advice during the early stages of this work, and Larry Abbott, Steve Colburn and Gabriel Soto for discussions.

7

References

Attias H and Schreiner CE. Coding of Naturalistic Stimuli by Auditory Midbrain Neurons. In: Advances in Neural Information Processing 10. Cambridge, MA: MIT Press, 1998, p. 103-109.Attneave F. Some information aspects of visual perception. Psychol Rev 61: 183-193, 1954.Barlow HB. Possible principles underlying the transformation of sensory messages. In: Sensory communication, edited by Rosenblith WA. Cambridge, MA: MIT Press, 1961, p. 217-234.

Berry MJ, 2nd and Meister M. Refractoriness and neural precision. J Neurosci 18: 2200-2211, 1998.Brainard MS and Doupe AJ. What songbirds teach us about learning. Nature 417: 351-358, 2002.Cruikshank SJ, Rose HJ, and Metherate R. Auditory thalamocortical synaptic transmission in vitro. J Neurophysiol 87: 361-384, 2002.

Dayan P and Abbott LF. Theoretical neuroscience: computational and mathematical modeling of neural systems. Cambridge, MA: MIT Press, 2001.

DeAngelis GC, Ohzawa I, and Freeman RD. Receptive-field dynamics in the central visual pathways [see comments]. Trends Neurosci 18: 451-458, 1995.

Depireux DA, Simon JZ, Klein DJ, and Shamma SA. Spectro-temporal response field characterization with dynamic ripples in ferret primary auditory cortex. J Neurophysiol 85: 1220-1234, 2001.

DeWeese MR, Wehr M, and Zador AM. Binary spiking in auditory cortex. J Neurosci 23: 7940-7949, 2003.

Edeline JM and Weinberger NM. Receptive field plasticity in the auditory cortex during frequency discrimination training: selective retuning independent of task difficulty. Behav Neurosci 107: 82-103, 1993.

Elhilali M, Fritz JB, Klein DJ, Simon JZ, and Shamma SA. Dynamics of precise spike timing in primary auditory cortex. J Neurosci 24: 1159-1172, 2004.

Fitch RH, Miller S, and Tallal P. Neurobiology of speech perception. Annu Rev Neurosci 20: 331-353, 1997.

Fritz J, Shamma S, Elhilali M, and Klein D. Rapid task-related plasticity of spectrotemporal receptive fields in primary auditory cortex. Nat Neurosci 6: 1216-1223, 2003.

Gentner TQ. Neural systems for individual song recognition in adult birds. Ann N Y Acad Sci 1016: 282-302, 2004.

Gentner TQ and Margoliash D. Neuronal populations and single cells representing learned auditory objects. Nature 424: 669-674, 2003.

Ghazanfar AA and Nicolelis MA. Feature article: the structure and function of dynamic cortical and thalamic receptive fields. Cereb Cortex 11: 183-193., 2001.

Grace JA, Amin N, Singh NC, and Theunissen FE. Selectivity for conspecific song in the zebra finch auditory forebrain. J Neurophysiol 89: 472-487, 2003.

Heil P. Auditory cortical onset responses revisited. I. First-spike timing. J Neurophysiol 77: 2616-2641, 1997.

Hsu A, Woolley SM, Fremouw TE, and Theunissen FE. Modulation power and phase spectrum of natural sounds enhance neural encoding performed by single auditory neurons. J Neurosci 24: 9201-9211, 2004.

Konishi M. Birdsong: from behavior to neuron. Annu Rev Neurosci 8: 125-170, 1985.Lewicki MS. Efficient coding of natural sounds. Nat Neurosci 5: 356-363, 2002.

Linden JF, Liu RC, Sahani M, Schreiner CE, and Merzenich MM. Spectrotemporal structure of

receptive fields in areas AI and AAF of mouse auditory cortex. J Neurophysiol 90: 2660-2675, 2003.Lu T, Liang L, and Wang X. Temporal and rate representations of time-varying signals in the auditory cortex of awake primates. Nat Neurosci 4: 1131-1138, 2001.

Machens CK, Schutze H, Franz A, Kolesnikova O, Stemmler MB, Ronacher B, and Herz AV. Single auditory neurons rapidly discriminate conspecific communication signals. Nat Neurosci 6: 341-342, 2003.

Machens CK, Wehr MS, and Zador AM. Linearity of cortical receptive fields measured with natural sounds. J Neurosci 24: 1089-1100, 2004.

8

Mello CV, Velho TA, and Pinaud R. Song-induced gene expression: a window on song auditory processing and perception. Ann N Y Acad Sci 1016: 263-281, 2004.

Mello CV, Vicario DS, and Clayton DF. Song presentation induces gene expression in the songbird forebrain. Proc Natl Acad Sci U S A 89: 6818-6822, 1992.

Miller LM, Escabi MA, Read HL, and Schreiner CE. Functional convergence of response properties in the auditory thalamocortical system. Neuron 32: 151-160., 2001.

Miller LM, Escabi MA, Read HL, and Schreiner CE. Spectrotemporal receptive fields in the lemniscal auditory thalamus and cortex. J Neurophysiol 87: 516-527., 2002.

Muller CM and Scheich H. Contribution of GABAergic inhibition to the response characteristics of auditory units in the avian forebrain. J Neurophysiol 59: 1673-1689, 1988.

Muller CM and Scheich H. GABAergic inhibition increases the neuronal selectivity to natural sounds in the avian auditory forebrain. Brain Res 414: 376-380, 1987.

Olshausen BA and Field DJ. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature 381: 607-609, 1996.

Qiu A, Schreiner CE, and Escabi MA. Gabor analysis of auditory midbrain receptive fields: spectro-temporal and binaural composition. J Neurophysiol 90: 456-476, 2003.

Rauschecker JP. Cortical processing of complex sounds. Curr Opin Neurobiol 8: 516-521, 1998.Recanzone GH, Schreiner CE, and Merzenich MM. Plasticity in the frequency representation of primary auditory cortex following discrimination training in adult owl monkeys. J Neurosci 13: 87-103, 1993.

Rieke F, Bodnar DA, and Bialek W. Naturalistic stimuli increase the rate and efficiency of information transmission by primary auditory afferents. Proc R Soc Lond B Biol Sci 262: 259-265, 1995.

Sen K, Theunissen FE, and Doupe AJ. Feature analysis of natural sounds in the songbird auditory forebrain. J Neurophysiol 86: 1445-1458., 2001.

Shannon RV, Zeng FG, Kamath V, Wygonski J, and Ekelid M. Speech recognition with primarily temporal cues. Science 270: 303-304, 1995.

Simoncelli EP and Olshausen BA. Natural image statistics and neural representation. Annu Rev Neurosci 24: 1193-1216, 2001.

Singh NC and Theunissen FE. Modulation spectra of natural sounds and ethological theories of auditory processing. J Acoust Soc Am 114: 3394-3411, 2003.

Tan AY, Zhang LI, Merzenich MM, and Schreiner CE. Tone-evoked excitatory and inhibitory synaptic conductances of primary auditory cortex neurons. J Neurophysiol 92: 630-643, 2004.

Theunissen FE, Amin N, Shaevitz SS, Woolley SM, Fremouw T, and Hauber ME. Song selectivity in the song system and in the auditory forebrain. Ann N Y Acad Sci 1016: 222-245, 2004.

Theunissen FE, Sen K, and Doupe AJ. Spectral-temporal receptive fields of nonlinear auditory neurons obtained using natural sounds. J Neurosci 20: 2315-2331, 2000.van Rossum MC. A novel spike distance. Neural Comput 13: 751-763, 2001.

Victor JD and Purpura KP. Metric-space analysis of spike-trains: theory, algorithms and application.Network: Computation in Neural Systems 8: 127-164, 1997.

Vinje WE and Gallant JL. Sparse coding and decorrelation in primary visual cortex during natural vision. Science 287: 1273-1276, 2000.

Wehr M and Zador AM. Balanced inhibition underlies tuning and sharpens spike timing in auditory cortex. Nature 426: 442-446, 2003.

Wehr M and Zador AM. Synaptic mechanisms of forward masking in rat auditory cortex. Computational and Systems Neuroscience Meeting Abs 280, 2005.

Woolley SM and Theunissen FE. Adaptive temporal processing and the precise encoding of

vocalizations in the songbird auditory midbrain. Association for Research in Otolaryngology Abs 978, 2005.

9

FigureLegends

Figure 1: Dynamics of song discrimination

a, The accuracy of song discrimination plotted as a function of spike train length at a timescale 󰁗= 10 ms. The spike trains were generated using a delayed inhibitory STRF (inset) and a stochastic spike-generator. The level of accuracy was computed using a supervised classification scheme. The chance level for the classification was 5% (gray dashed line). A similar result was obtained (black dashed line) using a cost-based distance metric, cost = 1/11 ms-1 (Victor and Purpura 1997). The plateau in the discrimination curve occurs due to the limited lengths of songs in our stimulus ensemble (see Methods). Truncating songs earlier resulted in a correspondingly earlier plateau (data not shown).b,The change in the level of discrimination for various choices of the timescale parameter 󰁗. The spike train duration was fixed at 1000 ms. The gray region represents one standard deviation about the mean value computed from 100 repetitions of the classification procedure.

Figure 2: The effect of delayed inhibition on song discrimination.

A comparison of the discrimination curves for a model with a purely excitatory STRF (right inset) and a model with a delayed inhibitory STRF (left inset). The spectral and temporal ranges of the STRFs are the same as in Fig. 1a. The timescale 󰁗=10 ms. The mean firing rates of both models, averaged over all songs, were constrained to be equal (see Methods). The model with the inhibitory subregion (circles) performs faster and more accurately than the model without inhibition (triangles).

Figure 3: A candidate microcircuit

a, A schematic of the circuit. The excitatory (E) input is generated from an STRF-based Poisson model (see Methods). The inhibitory input (I) is derived from the excitatory spike train by adding a delay and introducing a jitter to the spikes of the E input neuron. The output neuron is an integrate and fire neuron with excitatory and inhibitory synapses. The weights weand wicontrol the relative strengths of the two inputs. b, A comparison of the discrimination curves for output E neuron, with inhibition (circles) and without inhibition in the circuit (triangles). The timescale 󰁗=20 ms. The mean firing rate of the output neuron, averaged over all songs, was constrained to be equal in both models (see Methods). The presence of inhibition improves both the speed and accuracy of discrimination.

10

Figure 1

11

Figure 2

12

Figure 3

13