Central auditory neurons have composite receptive fields

Abstract

High-level neurons processing complex, behaviorally relevant signals are sensitive to conjunctions of features. Characterizing the receptive fields of such neurons is difficult with standard statistical tools, however, and the principles governing their organization remain poorly understood. Here, we demonstrate multiple distinct receptive-field features in individual high-level auditory neurons in a songbird, European starling, in response to natural vocal signals (songs). We then show that receptive fields with similar characteristics can be reproduced by an unsupervised neural network trained to represent starling songs with a single learning rule that enforces sparseness and divisive normalization. We conclude that central auditory neurons have composite receptive fields that can arise through a combination of sparseness and normalization in neural circuits. Our results, along with descriptions of random, discontinuous receptive fields in the central olfactory neurons in mammals and insects, suggest general principles of neural computation across sensory systems and animal classes.

Keywords: auditory system; neural networks; receptive fields; sparseness; unsupervised learning.

Fig. 1.

A single NCM neuron responds to several different motifs. (A) Spike raster plot (Top), peri-stimulus histogram (Middle), and spectrogram (Bottom) showing an example NCM neuron’s response to a full song. (B) Four 3-second-long excerpts taken at the indicated times from the response shown in A. The different panels show the neuron responding to acoustically distinct motifs (harmonic stacks, clicks, and other broadband stimuli), supporting the idea that individual NCM neurons can be sensitive to a variety of different stimulus features.

Fig. 2.

Composite receptive field of a single NCM neuron. (A) Examples of multiple excitatory and suppressive features obtained from one NCM neuron. The top two rows show the negative (excitatory) features. In this neuron, eight negative eigenvalues were significant. The largest nonsignificant eigenvector is also displayed to the right of the dotted line in row two and can be seen to contain structure. Eigenvectors corresponding to even smaller eigenvalues (not shown) contained no clear structure. The two bottom rows show nine significant positive (suppressive) features. (B) Eigenspectrum of the matrix J for the same neuron as in A. Eigenvalues were normalized for comparison with the data in Fig. 6B. The dashed lines indicate the two largest (in absolute value) positive and negative eigenvalues obtained from 500 symmetrical Gaussian random matrices with the same mean and variance as those of matrix J.

Fig. 3.

Capturing the statistics of feature ensembles. Projections of modulation power spectra for starling songs (green), MNE features (red), and artificial neural network features (black) on spectral (Left) and temporal (Right) axes.

Fig. 4.

Prediction of responses to new stimuli. (A) The empirically measured time-varying average spike rate (blue) and the MNE-predicted spike rate (red) for a single neuron’s response to a song. The correlation coefficient between the measured and predicted response was 0.56. The number of spikes in each time bin was normalized by the number of stimulus repetitions. (B) Full distribution of correlation coefficients obtained with the second-order MNE model plotted against those obtained with the STRF. The diagonal line indicates unity.

Fig. 5.

Composite receptive fields from larger datasets. (A) Examples of excitatory and suppressive features obtained from an NCM neuron using 60 1-minute-long songs. The top two rows show the nine significant excitatory features, and the two bottom rows show seven significant suppressive features. Two largest nonsignificant eigenvectors are also displayed (right-hand side of dotted line) and can be seen to contain structure and not only noise. (B) Eigenspectrum of the matrix J for the neuron in A. Eigenvalues were normalized for comparison with the data in Fig. 6B. The dashed lines indicate the two largest (in absolute value) positive and negative eigenvalues obtained from 500 symmetrical Gaussian random matrices with the same mean and variance as those of J.

Fig. 6.

Neural network trained on starling songs learns composite receptive fields. (A) Ten most active features for seven randomly chosen layer-2 units (a to g). (B) A histogram showing a distribution of the basis features’ activities for one layer-2 unit. The absolute normalized magnitude is shown for comparison with the distributions of features’ magnitudes in Figs. 2B and 5B. Note that most features are close to zero (lifetime sparseness) and that the most active features are of the same order of magnitude, as expected. (C) Cumulative density function (CDF) showing the percentage difference in the pair-wise magnitude between neighboring most active features of all layer-2 units. Ten largest basis features are selected for each layer-2 unit and sorted according to their activity (magnitude), then the percent difference is taken between the neighboring values.