Limits...
Neuronal precision and the limits for acoustic signal recognition in a small neuronal network.

Neuhofer D, Stemmler M, Ronacher B - J. Comp. Physiol. A Neuroethol. Sens. Neural. Behav. Physiol. (2010)

Bottom Line: By progressively corrupting the envelope of a female song, we determined the critical degradation level at which males failed to recognize a courtship call in behavioral experiments.At consecutive levels of processing, intrinsic variability increased, while the sensitivity to external noise decreased.We followed two approaches to determine critical degradation levels from spike train dissimilarities, and compared the results with the limits of signal recognition measured in behaving animals.

View Article: PubMed Central - PubMed

Affiliation: Department of Biology, Humboldt-Universität zu Berlin, Invalidenstrasse 43, 10115, Berlin, Germany. neuhofda@cms.hu-berlin.de

ABSTRACT
Recognition of acoustic signals may be impeded by two factors: extrinsic noise, which degrades sounds before they arrive at the receiver's ears, and intrinsic neuronal noise, which reveals itself in the trial-to-trial variability of the responses to identical sounds. Here we analyzed how these two noise sources affect the recognition of acoustic signals from potential mates in grasshoppers. By progressively corrupting the envelope of a female song, we determined the critical degradation level at which males failed to recognize a courtship call in behavioral experiments. Using the same stimuli, we recorded intracellularly from auditory neurons at three different processing levels, and quantified the corresponding changes in spike train patterns by a spike train metric, which assigns a distance between spike trains. Unexpectedly, for most neurons, intrinsic variability accounted for the main part of the metric distance between spike trains, even at the strongest degradation levels. At consecutive levels of processing, intrinsic variability increased, while the sensitivity to external noise decreased. We followed two approaches to determine critical degradation levels from spike train dissimilarities, and compared the results with the limits of signal recognition measured in behaving animals.

Show MeSH

Related in: MedlinePlus

Two approaches to determine a neuronal critical degradation level (nCDL). a Distance matrix of the spike trains of a local neuron, SN3. Metric distances are color coded from blue to red (zero to maximum distance). b Cumulative nCDLs as revealed by the discrimination analysis. The colored curves (see inset) indicate the percentage of cells (ordinate) for which the distances x0 and yn (abscissa) could be discriminated with an accuracy rate of at least 95% (closed circles, solid lines) or 97% (open circles, dashed lines). For receptor cells there was no difference between 95 and 97%. Different processing stages (receptors, local and ascending neurons) are shown separately. Behavioral CDLs are drawn in gray (bCDL data from Fig. 4). c Result of the information-based clustering algorithm for the spike trains of SN3 recording. Along the ordinate, the graph displays the mean probability that the responses to a distinct degradation level (abscissa) belong to the ‘orig’ class (containing the spike trains in response to the original song) or to the ‘noise’ class (containing the spike trains in response to the remaining degradation levels). The arrow indicates the neuronal critical degradation level (nCDL). d Frequency of assignments to the ‘noise’ class (ordinate) in dependence of different degradation levels (abscissa). Different levels of computation are drawn separately. Sampling size for b and c: REC = 7, LN = 22, AN = 19
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC3040818&req=5

Fig5: Two approaches to determine a neuronal critical degradation level (nCDL). a Distance matrix of the spike trains of a local neuron, SN3. Metric distances are color coded from blue to red (zero to maximum distance). b Cumulative nCDLs as revealed by the discrimination analysis. The colored curves (see inset) indicate the percentage of cells (ordinate) for which the distances x0 and yn (abscissa) could be discriminated with an accuracy rate of at least 95% (closed circles, solid lines) or 97% (open circles, dashed lines). For receptor cells there was no difference between 95 and 97%. Different processing stages (receptors, local and ascending neurons) are shown separately. Behavioral CDLs are drawn in gray (bCDL data from Fig. 4). c Result of the information-based clustering algorithm for the spike trains of SN3 recording. Along the ordinate, the graph displays the mean probability that the responses to a distinct degradation level (abscissa) belong to the ‘orig’ class (containing the spike trains in response to the original song) or to the ‘noise’ class (containing the spike trains in response to the remaining degradation levels). The arrow indicates the neuronal critical degradation level (nCDL). d Frequency of assignments to the ‘noise’ class (ordinate) in dependence of different degradation levels (abscissa). Different levels of computation are drawn separately. Sampling size for b and c: REC = 7, LN = 22, AN = 19

Mentions: Neurons differed in their robustness to noise, typically maintaining the key features of the spike train patterns up to a certain degradation level. Even in response to the original song, quite some trial-to-trial variability was observed, which is reflected in the set of ‘intrinsic’ distances between the spike trains. As the level of signal degradation increases, a point will be reached at which the corresponding spike train distances will differ significantly from these ‘intrinsic’ distances. At this point, the overlap between the distance distributions, taken from the outer right column of the distance matrix (see blue frame in Fig. 5a), falls below a fixed fraction. This fraction, typically 5%, defines threshold discriminability. Figure 5b shows the percentage of cells whose spike train patterns were distinguishable with p = 0.05 from the spike trains in response to the original stimulus, plotted as a function of degradation level. The number of receptor cells whose spike trains surpass threshold discriminability increased steeply between −9 and −3 dB and then leveled off. The local neurons showed the strongest increase up to 0 dB. The behavioral data (bCDL 50%) also exhibits a steep increase of cumulative bCDL between −9 and 0 dB. In contrast, fewer than 35% of the ascending neurons exhibited a significant difference even at the highest degradation level (Fig. 5b, see also Fig. 2a). Figure 5a shows the distance matrix of a local interneuron (SN3), which did not achieve the strict criterion of 95% discrimination performance. Nevertheless, the distance matrix of this cell shows a distinct block-like structure, such that the distances between spike trains in response to −3 to 9 dB degradation group together. To take advantage of this distinct response property of many neurons, we applied a second, more sensitive classification procedure to determine critical classification boundaries, going beyond solely using the outer right-hand column of the distance matrix. The unsupervised information-clustering algorithm (Slonim et al. 2005) uses the data of the whole matrix (see stippled frame in Fig. 5a) to cluster the spike train distances into two distinct classes. Figure 5c plots the mean probability, as determined by the clustering algorithm, that the spike trains of this SN3 neuron belong to the class of the uncorrupted stimulus (closed circles) or the ‘noise’ class (open circles). For these recordings, the mean probability of a spike train belonging to the first class dropped to 50% between −6 and −3 dB. Spike trains in response to the original song and the first and second degradation level were assigned to the ‘orig’ class, spike trains in response to higher degradation levels were assigned to the ‘noise’ class. For the local interneuron (SN3) in Fig. 5a and c, the intrinsic and extrinsic distance distributions overlapped considerably, even at the highest noise level, and simply comparing the distances failed to reach the 95% criterion in this case. Only the information-clustering algorithm managed to classify the spike trains reliably. Other examples that exhibited a simple boundary between the two classes are shown in Online Resource 6. But there were also cells that exhibited ambiguous classification boundaries. Examples of such cells are shown in Online Resource 7. We performed a conservative analysis, for which stringent requirements were applied to eliminate cells with ambiguous boundaries (see “Materials and methods”). Figure 5d shows the percentage of neurons whose responses were assigned to the ‘noise’ class as a function of the degradation level. Between −6 and 3 dB there was a steep increase in the number of receptor cells; the same held true for ascending neurons. For local neurons, this steep increase was shifted to 3 dB higher degradation levels.Fig. 5


Neuronal precision and the limits for acoustic signal recognition in a small neuronal network.

Neuhofer D, Stemmler M, Ronacher B - J. Comp. Physiol. A Neuroethol. Sens. Neural. Behav. Physiol. (2010)

Two approaches to determine a neuronal critical degradation level (nCDL). a Distance matrix of the spike trains of a local neuron, SN3. Metric distances are color coded from blue to red (zero to maximum distance). b Cumulative nCDLs as revealed by the discrimination analysis. The colored curves (see inset) indicate the percentage of cells (ordinate) for which the distances x0 and yn (abscissa) could be discriminated with an accuracy rate of at least 95% (closed circles, solid lines) or 97% (open circles, dashed lines). For receptor cells there was no difference between 95 and 97%. Different processing stages (receptors, local and ascending neurons) are shown separately. Behavioral CDLs are drawn in gray (bCDL data from Fig. 4). c Result of the information-based clustering algorithm for the spike trains of SN3 recording. Along the ordinate, the graph displays the mean probability that the responses to a distinct degradation level (abscissa) belong to the ‘orig’ class (containing the spike trains in response to the original song) or to the ‘noise’ class (containing the spike trains in response to the remaining degradation levels). The arrow indicates the neuronal critical degradation level (nCDL). d Frequency of assignments to the ‘noise’ class (ordinate) in dependence of different degradation levels (abscissa). Different levels of computation are drawn separately. Sampling size for b and c: REC = 7, LN = 22, AN = 19
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC3040818&req=5

Fig5: Two approaches to determine a neuronal critical degradation level (nCDL). a Distance matrix of the spike trains of a local neuron, SN3. Metric distances are color coded from blue to red (zero to maximum distance). b Cumulative nCDLs as revealed by the discrimination analysis. The colored curves (see inset) indicate the percentage of cells (ordinate) for which the distances x0 and yn (abscissa) could be discriminated with an accuracy rate of at least 95% (closed circles, solid lines) or 97% (open circles, dashed lines). For receptor cells there was no difference between 95 and 97%. Different processing stages (receptors, local and ascending neurons) are shown separately. Behavioral CDLs are drawn in gray (bCDL data from Fig. 4). c Result of the information-based clustering algorithm for the spike trains of SN3 recording. Along the ordinate, the graph displays the mean probability that the responses to a distinct degradation level (abscissa) belong to the ‘orig’ class (containing the spike trains in response to the original song) or to the ‘noise’ class (containing the spike trains in response to the remaining degradation levels). The arrow indicates the neuronal critical degradation level (nCDL). d Frequency of assignments to the ‘noise’ class (ordinate) in dependence of different degradation levels (abscissa). Different levels of computation are drawn separately. Sampling size for b and c: REC = 7, LN = 22, AN = 19
Mentions: Neurons differed in their robustness to noise, typically maintaining the key features of the spike train patterns up to a certain degradation level. Even in response to the original song, quite some trial-to-trial variability was observed, which is reflected in the set of ‘intrinsic’ distances between the spike trains. As the level of signal degradation increases, a point will be reached at which the corresponding spike train distances will differ significantly from these ‘intrinsic’ distances. At this point, the overlap between the distance distributions, taken from the outer right column of the distance matrix (see blue frame in Fig. 5a), falls below a fixed fraction. This fraction, typically 5%, defines threshold discriminability. Figure 5b shows the percentage of cells whose spike train patterns were distinguishable with p = 0.05 from the spike trains in response to the original stimulus, plotted as a function of degradation level. The number of receptor cells whose spike trains surpass threshold discriminability increased steeply between −9 and −3 dB and then leveled off. The local neurons showed the strongest increase up to 0 dB. The behavioral data (bCDL 50%) also exhibits a steep increase of cumulative bCDL between −9 and 0 dB. In contrast, fewer than 35% of the ascending neurons exhibited a significant difference even at the highest degradation level (Fig. 5b, see also Fig. 2a). Figure 5a shows the distance matrix of a local interneuron (SN3), which did not achieve the strict criterion of 95% discrimination performance. Nevertheless, the distance matrix of this cell shows a distinct block-like structure, such that the distances between spike trains in response to −3 to 9 dB degradation group together. To take advantage of this distinct response property of many neurons, we applied a second, more sensitive classification procedure to determine critical classification boundaries, going beyond solely using the outer right-hand column of the distance matrix. The unsupervised information-clustering algorithm (Slonim et al. 2005) uses the data of the whole matrix (see stippled frame in Fig. 5a) to cluster the spike train distances into two distinct classes. Figure 5c plots the mean probability, as determined by the clustering algorithm, that the spike trains of this SN3 neuron belong to the class of the uncorrupted stimulus (closed circles) or the ‘noise’ class (open circles). For these recordings, the mean probability of a spike train belonging to the first class dropped to 50% between −6 and −3 dB. Spike trains in response to the original song and the first and second degradation level were assigned to the ‘orig’ class, spike trains in response to higher degradation levels were assigned to the ‘noise’ class. For the local interneuron (SN3) in Fig. 5a and c, the intrinsic and extrinsic distance distributions overlapped considerably, even at the highest noise level, and simply comparing the distances failed to reach the 95% criterion in this case. Only the information-clustering algorithm managed to classify the spike trains reliably. Other examples that exhibited a simple boundary between the two classes are shown in Online Resource 6. But there were also cells that exhibited ambiguous classification boundaries. Examples of such cells are shown in Online Resource 7. We performed a conservative analysis, for which stringent requirements were applied to eliminate cells with ambiguous boundaries (see “Materials and methods”). Figure 5d shows the percentage of neurons whose responses were assigned to the ‘noise’ class as a function of the degradation level. Between −6 and 3 dB there was a steep increase in the number of receptor cells; the same held true for ascending neurons. For local neurons, this steep increase was shifted to 3 dB higher degradation levels.Fig. 5

Bottom Line: By progressively corrupting the envelope of a female song, we determined the critical degradation level at which males failed to recognize a courtship call in behavioral experiments.At consecutive levels of processing, intrinsic variability increased, while the sensitivity to external noise decreased.We followed two approaches to determine critical degradation levels from spike train dissimilarities, and compared the results with the limits of signal recognition measured in behaving animals.

View Article: PubMed Central - PubMed

Affiliation: Department of Biology, Humboldt-Universität zu Berlin, Invalidenstrasse 43, 10115, Berlin, Germany. neuhofda@cms.hu-berlin.de

ABSTRACT
Recognition of acoustic signals may be impeded by two factors: extrinsic noise, which degrades sounds before they arrive at the receiver's ears, and intrinsic neuronal noise, which reveals itself in the trial-to-trial variability of the responses to identical sounds. Here we analyzed how these two noise sources affect the recognition of acoustic signals from potential mates in grasshoppers. By progressively corrupting the envelope of a female song, we determined the critical degradation level at which males failed to recognize a courtship call in behavioral experiments. Using the same stimuli, we recorded intracellularly from auditory neurons at three different processing levels, and quantified the corresponding changes in spike train patterns by a spike train metric, which assigns a distance between spike trains. Unexpectedly, for most neurons, intrinsic variability accounted for the main part of the metric distance between spike trains, even at the strongest degradation levels. At consecutive levels of processing, intrinsic variability increased, while the sensitivity to external noise decreased. We followed two approaches to determine critical degradation levels from spike train dissimilarities, and compared the results with the limits of signal recognition measured in behaving animals.

Show MeSH
Related in: MedlinePlus