Limits...
Discriminative learning of receptive fields from responses to non-Gaussian stimulus ensembles.

Meyer AF, Diepenbrock JP, Happel MF, Ohl FW, Anemüller J - PLoS ONE (2014)

Bottom Line: Results demonstrate that the proposed approach successfully identifies correct underlying STRFs, even in cases where second-order methods based on the spike-triggered average (STA) do not.Applied to small data samples, the method is shown to converge on smaller amounts of experimental recordings and with lower estimation variance than the generalized linear model and recent information theoretic methods.Thus, CbRF estimation may prove useful for investigation of neuronal processes in response to natural stimuli and in settings where rapid adaptation is induced by experimental design.

View Article: PubMed Central - PubMed

Affiliation: Department of Medical Physics and Acoustics and Cluster of Excellence ''Hearing4all'', University of Oldenburg, Oldenburg, Germany.

ABSTRACT
Analysis of sensory neurons' processing characteristics requires simultaneous measurement of presented stimuli and concurrent spike responses. The functional transformation from high-dimensional stimulus space to the binary space of spike and non-spike responses is commonly described with linear-nonlinear models, whose linear filter component describes the neuron's receptive field. From a machine learning perspective, this corresponds to the binary classification problem of discriminating spike-eliciting from non-spike-eliciting stimulus examples. The classification-based receptive field (CbRF) estimation method proposed here adapts a linear large-margin classifier to optimally predict experimental stimulus-response data and subsequently interprets learned classifier weights as the neuron's receptive field filter. Computational learning theory provides a theoretical framework for learning from data and guarantees optimality in the sense that the risk of erroneously assigning a spike-eliciting stimulus example to the non-spike class (and vice versa) is minimized. Efficacy of the CbRF method is validated with simulations and for auditory spectro-temporal receptive field (STRF) estimation from experimental recordings in the auditory midbrain of Mongolian gerbils. Acoustic stimulation is performed with frequency-modulated tone complexes that mimic properties of natural stimuli, specifically non-Gaussian amplitude distribution and higher-order correlations. Results demonstrate that the proposed approach successfully identifies correct underlying STRFs, even in cases where second-order methods based on the spike-triggered average (STA) do not. Applied to small data samples, the method is shown to converge on smaller amounts of experimental recordings and with lower estimation variance than the generalized linear model and recent information theoretic methods. Thus, CbRF estimation may prove useful for investigation of neuronal processes in response to natural stimuli and in settings where rapid adaptation is induced by experimental design.

Show MeSH

Related in: MedlinePlus

Spectro-temporal receptive field (STRF) estimation from simulated responses to natural stimuli: Robustness to neuronal nonlinearity.(A) Ground truth spectro-temporal linear RF filter used in LNP model simulations of spike responses to four minutes of human speech. (B) Different static nonlinearities utilized in the LNP model, ranging from linear to step-like, the output of which was used for Poisson process spike train generation. (C) Linear RF filter estimates obtained with four estimation methods (rows, explanation cf. Table 1) for each of the nonlinearities in panel B (columns). Numbers indicate correlation of estimated with true RF filter. CbRF and MID methods reliably recovered the true linear filters. The GLM shows a bias when the assumed exponential inverse link function deviates from the static nonlinearity used to generate the data, e.g., for the compressive, sigmoid, and threshold nonlinearities. (D) Average correlation between true and estimated linear filter for speech stimuli of varying length. An ensemble of model cells was created using different linear filters and different nonlinearities from panel B with randomly chosen parameters. Shown are the correlations' mean and standard deviation across 150 model cells for each method. With mean correlation about 0.93 for 100% (four minutes) of the data, CbRF and MID yield higher correlation than GLM and ridge regression. Towards smaller sample sizes, CbRF method performance declines slower than the other methods' including MID's. Bias of the linear ridge regression estimator may be due to the highly non-Gaussian structure of human speech. (E) Same experiment as in D but with conspecific zebra finch vocalization stimuli of total length three minutes. CbRF method resulted in highest mean correlation for all stimuli lengths. GLM and MID method showed similar performance for long stimuli with GLM declining less towards smaller sample sizes below 50%. The somewhat higher mean correlation values observed for ridge regression in comparison to panel D may be attributed to the fact that the zebra finch vocalizations were less non-Gaussian than human speech.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3974709&req=5

pone-0093062-g006: Spectro-temporal receptive field (STRF) estimation from simulated responses to natural stimuli: Robustness to neuronal nonlinearity.(A) Ground truth spectro-temporal linear RF filter used in LNP model simulations of spike responses to four minutes of human speech. (B) Different static nonlinearities utilized in the LNP model, ranging from linear to step-like, the output of which was used for Poisson process spike train generation. (C) Linear RF filter estimates obtained with four estimation methods (rows, explanation cf. Table 1) for each of the nonlinearities in panel B (columns). Numbers indicate correlation of estimated with true RF filter. CbRF and MID methods reliably recovered the true linear filters. The GLM shows a bias when the assumed exponential inverse link function deviates from the static nonlinearity used to generate the data, e.g., for the compressive, sigmoid, and threshold nonlinearities. (D) Average correlation between true and estimated linear filter for speech stimuli of varying length. An ensemble of model cells was created using different linear filters and different nonlinearities from panel B with randomly chosen parameters. Shown are the correlations' mean and standard deviation across 150 model cells for each method. With mean correlation about 0.93 for 100% (four minutes) of the data, CbRF and MID yield higher correlation than GLM and ridge regression. Towards smaller sample sizes, CbRF method performance declines slower than the other methods' including MID's. Bias of the linear ridge regression estimator may be due to the highly non-Gaussian structure of human speech. (E) Same experiment as in D but with conspecific zebra finch vocalization stimuli of total length three minutes. CbRF method resulted in highest mean correlation for all stimuli lengths. GLM and MID method showed similar performance for long stimuli with GLM declining less towards smaller sample sizes below 50%. The somewhat higher mean correlation values observed for ridge regression in comparison to panel D may be attributed to the fact that the zebra finch vocalizations were less non-Gaussian than human speech.

Mentions: Responses were simulated using a narrow-band onset spectro-temporal receptive field (STRF) (Figure 6A), a pattern that has been found throughout different stages of the auditory system of mammals [10], [15], [45]. The output of the linear stage was transformed into a spike rate using different nonlinearities, ranging from linear to step-like (cf. Figure 6B). Spikes were generated from the spike rate by an inhomogeneous Poisson process. We strove to achieve a realistic average spike rate between 0.02 and 0.1 spikes per sample for all nonlinearities.


Discriminative learning of receptive fields from responses to non-Gaussian stimulus ensembles.

Meyer AF, Diepenbrock JP, Happel MF, Ohl FW, Anemüller J - PLoS ONE (2014)

Spectro-temporal receptive field (STRF) estimation from simulated responses to natural stimuli: Robustness to neuronal nonlinearity.(A) Ground truth spectro-temporal linear RF filter used in LNP model simulations of spike responses to four minutes of human speech. (B) Different static nonlinearities utilized in the LNP model, ranging from linear to step-like, the output of which was used for Poisson process spike train generation. (C) Linear RF filter estimates obtained with four estimation methods (rows, explanation cf. Table 1) for each of the nonlinearities in panel B (columns). Numbers indicate correlation of estimated with true RF filter. CbRF and MID methods reliably recovered the true linear filters. The GLM shows a bias when the assumed exponential inverse link function deviates from the static nonlinearity used to generate the data, e.g., for the compressive, sigmoid, and threshold nonlinearities. (D) Average correlation between true and estimated linear filter for speech stimuli of varying length. An ensemble of model cells was created using different linear filters and different nonlinearities from panel B with randomly chosen parameters. Shown are the correlations' mean and standard deviation across 150 model cells for each method. With mean correlation about 0.93 for 100% (four minutes) of the data, CbRF and MID yield higher correlation than GLM and ridge regression. Towards smaller sample sizes, CbRF method performance declines slower than the other methods' including MID's. Bias of the linear ridge regression estimator may be due to the highly non-Gaussian structure of human speech. (E) Same experiment as in D but with conspecific zebra finch vocalization stimuli of total length three minutes. CbRF method resulted in highest mean correlation for all stimuli lengths. GLM and MID method showed similar performance for long stimuli with GLM declining less towards smaller sample sizes below 50%. The somewhat higher mean correlation values observed for ridge regression in comparison to panel D may be attributed to the fact that the zebra finch vocalizations were less non-Gaussian than human speech.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3974709&req=5

pone-0093062-g006: Spectro-temporal receptive field (STRF) estimation from simulated responses to natural stimuli: Robustness to neuronal nonlinearity.(A) Ground truth spectro-temporal linear RF filter used in LNP model simulations of spike responses to four minutes of human speech. (B) Different static nonlinearities utilized in the LNP model, ranging from linear to step-like, the output of which was used for Poisson process spike train generation. (C) Linear RF filter estimates obtained with four estimation methods (rows, explanation cf. Table 1) for each of the nonlinearities in panel B (columns). Numbers indicate correlation of estimated with true RF filter. CbRF and MID methods reliably recovered the true linear filters. The GLM shows a bias when the assumed exponential inverse link function deviates from the static nonlinearity used to generate the data, e.g., for the compressive, sigmoid, and threshold nonlinearities. (D) Average correlation between true and estimated linear filter for speech stimuli of varying length. An ensemble of model cells was created using different linear filters and different nonlinearities from panel B with randomly chosen parameters. Shown are the correlations' mean and standard deviation across 150 model cells for each method. With mean correlation about 0.93 for 100% (four minutes) of the data, CbRF and MID yield higher correlation than GLM and ridge regression. Towards smaller sample sizes, CbRF method performance declines slower than the other methods' including MID's. Bias of the linear ridge regression estimator may be due to the highly non-Gaussian structure of human speech. (E) Same experiment as in D but with conspecific zebra finch vocalization stimuli of total length three minutes. CbRF method resulted in highest mean correlation for all stimuli lengths. GLM and MID method showed similar performance for long stimuli with GLM declining less towards smaller sample sizes below 50%. The somewhat higher mean correlation values observed for ridge regression in comparison to panel D may be attributed to the fact that the zebra finch vocalizations were less non-Gaussian than human speech.
Mentions: Responses were simulated using a narrow-band onset spectro-temporal receptive field (STRF) (Figure 6A), a pattern that has been found throughout different stages of the auditory system of mammals [10], [15], [45]. The output of the linear stage was transformed into a spike rate using different nonlinearities, ranging from linear to step-like (cf. Figure 6B). Spikes were generated from the spike rate by an inhomogeneous Poisson process. We strove to achieve a realistic average spike rate between 0.02 and 0.1 spikes per sample for all nonlinearities.

Bottom Line: Results demonstrate that the proposed approach successfully identifies correct underlying STRFs, even in cases where second-order methods based on the spike-triggered average (STA) do not.Applied to small data samples, the method is shown to converge on smaller amounts of experimental recordings and with lower estimation variance than the generalized linear model and recent information theoretic methods.Thus, CbRF estimation may prove useful for investigation of neuronal processes in response to natural stimuli and in settings where rapid adaptation is induced by experimental design.

View Article: PubMed Central - PubMed

Affiliation: Department of Medical Physics and Acoustics and Cluster of Excellence ''Hearing4all'', University of Oldenburg, Oldenburg, Germany.

ABSTRACT
Analysis of sensory neurons' processing characteristics requires simultaneous measurement of presented stimuli and concurrent spike responses. The functional transformation from high-dimensional stimulus space to the binary space of spike and non-spike responses is commonly described with linear-nonlinear models, whose linear filter component describes the neuron's receptive field. From a machine learning perspective, this corresponds to the binary classification problem of discriminating spike-eliciting from non-spike-eliciting stimulus examples. The classification-based receptive field (CbRF) estimation method proposed here adapts a linear large-margin classifier to optimally predict experimental stimulus-response data and subsequently interprets learned classifier weights as the neuron's receptive field filter. Computational learning theory provides a theoretical framework for learning from data and guarantees optimality in the sense that the risk of erroneously assigning a spike-eliciting stimulus example to the non-spike class (and vice versa) is minimized. Efficacy of the CbRF method is validated with simulations and for auditory spectro-temporal receptive field (STRF) estimation from experimental recordings in the auditory midbrain of Mongolian gerbils. Acoustic stimulation is performed with frequency-modulated tone complexes that mimic properties of natural stimuli, specifically non-Gaussian amplitude distribution and higher-order correlations. Results demonstrate that the proposed approach successfully identifies correct underlying STRFs, even in cases where second-order methods based on the spike-triggered average (STA) do not. Applied to small data samples, the method is shown to converge on smaller amounts of experimental recordings and with lower estimation variance than the generalized linear model and recent information theoretic methods. Thus, CbRF estimation may prove useful for investigation of neuronal processes in response to natural stimuli and in settings where rapid adaptation is induced by experimental design.

Show MeSH
Related in: MedlinePlus