Stimulus encoding and feature extraction by multiple sensory neurons

Abstract

Neighboring cells in topographical sensory maps may transmit similar information to the next higher level of processing. How information transmission by groups of nearby neurons compares with the performance of single cells is a very important question for understanding the functioning of the nervous system. To tackle this problem, we quantified stimulus-encoding and feature extraction performance by pairs of simultaneously recorded electrosensory pyramidal cells in the hindbrain of weakly electric fish. These cells constitute the output neurons of the first central nervous stage of electrosensory processing. Using random amplitude modulations (RAMs) of a mimic of the fish's own electric field within behaviorally relevant frequency bands, we found that pyramidal cells with overlapping receptive fields exhibit strong stimulus-induced correlations. To quantify the encoding of the RAM time course, we estimated the stimuli from simultaneously recorded spike trains and found significant improvements over single spike trains. The quality of stimulus reconstruction, however, was still inferior to the one measured for single primary sensory afferents. In an analysis of feature extraction, we found that spikes of pyramidal cell pairs coinciding within a time window of a few milliseconds performed significantly better at detecting upstrokes and downstrokes of the stimulus compared with isolated spikes and even spike bursts of single cells. Coincident spikes can thus be considered "distributed bursts." Our results suggest that stimulus encoding by primary sensory afferents is transformed into feature extraction at the next processing stage. There, stimulus-induced coincident activity can improve the extraction of behaviorally relevant features from the stimulus.

Fig. 1.

Correlated activity of simultaneously recorded pyramidal cells. a, Representative raster plot segments of the spike trains of two simultaneously recorded I-units with overlapping receptive fields. The top trace shows the time course of the random amplitude modulation (cutoff frequency,f_c = 10 Hz; contrast, 25%). Action potentials occurring within a burst of spikes are indicated by thethick bar. The same stimulus was repeated five times, yielding five raster lines for each neuron. b, Cross-correlograms of the responses of the two I-units computed with a bin size of 3 msec. The x-axis indicates the time lag between the coincident spikes. The strong peak centered at 6.3 msec indicates that these two I-units fired coincident spikes within small time windows. The horizontal dashed line gives the expected value for two homogeneous Poisson neurons of the same firing rates as the recorded units firing independently. The peak and width (37 msec) of the responses are marked by vertical andhorizontal arrows, respectively. Inset, Shuffle-corrected cross-correlogram. The horizontal lineat 0 indicates the expected value for independent responses, and the dashed lines show the 2ς confidence limits under this null hypothesis (see Materials and Methods). Because the solid curve fell between these bounds, we conclude that the coincident activity is primarily stimulus induced. The average firing rates for the two units were 9.4 and 15.2 spikes/sec, respectively. c, Cross-correlogram of the responses of one E- and one I-unit. The center trough shows that these cells of opposing response type fired in anticorrelation. The minimum occurred at −0.2 msec; the width at half-height was 10 msec.Inset, Shuffle-corrected cross-correlogram. The average firing rates for the two units were 17.3 and 12.3 spikes/sec, respectively.

Fig. 2.

Properties of the cross-correlograms for pairs of units of the same type (n = 16). a, Distribution of the time lags at which the maximum occurred. Bin size, 5 msec.b, Distribution of the maxima of the cross-correlograms. Bin size, 0.25 coincidences (coinc/s). Thex-axis was cut at five coincidences/sec for clarity; there were three values beyond the axis limit (at 7.2, 9.3, and 19.1 coincidences/sec). c, Distribution of the widths at half-height of the peaks. Bin size, 25 msec. a–c,f_c = 5 Hz. For each neuronal pair, values for five stimulus contrasts are included.

Fig. 3.

Summarized results of stimulus estimation from spike trains of P-receptor afferents (P-aff.), single pyramidal (pyr.) cells, and pairs of simultaneously recorded pyramidal cells of the same type (E–E and I–I) and of opposite (opp.) type (E–I). The accuracy of the information transmitted about the time course of a stimulus is characterized by the coding fraction. Error bars indicate SD.f_c = 5 Hz. n, Overall number of experimental conditions (contrasts) for all cells or cell pairs analyzed. Data for P-receptor afferents taken from Kreiman et al. (2000).

Fig. 4.

Euclidian features and coincident spikes for the pair of I-type pyramidal cells depicted in Figure 1.a, Euclidian (Eucl.) feature for each of the two cells. b, Raster plot example highlighting those spikes that occur synchronously within a time window of ±5 msec asthick bars. c, The proportion of coincident spikes with respect to the total number of spikes for neuron A (top) and neuron B (bottom) is shown asblack bars as a function of the size of the coincidence window. The percentage of spikes that occur in bursts and coincide are shown as white bars. The overall percentage of bursting spikes is indicated as a gray bar at theright.

Fig. 5.

Feature extraction by the same pair of I-type pyramidal cells illustrated in Figure 1. a, Minimum probability of misclassification, p_E, by those spikes of neurons A and B, respectively, which had a coincident spike on the respective other neuron plotted against the size of the coincidence time window. p_E is the average of two error probabilities: in the case of this I-unit pair, these are the probability that coincident spikes are fired even when there is no downstroke in stimulus amplitude (false alarms) and the probability that a downstroke occurs but fails to elicit coincident spikes (misses). p_E decreased with decreasing size of the coincidence time window, indicating that spikes coinciding within a time window of ±5 msec transmit the information on the occurrence of stimulus features more reliably than spikes of single neurons. Filled symbols, Neuron A; open symbols, neuron B. b, Single-neuron performance of units A and B, respectively. isol., Isolated.

Fig. 6.

Summary diagram of feature extraction performance by ELL pyramidal cells. From left toright, bars indicate the averagep_E for coincident spikes of E–E pairs and I–I pairs, for coincident spikes of E–E and I–I pairs after shuffling of trials, for spike bursts of single E- and single I-units, for isolated spikes of single E- and I-units, and for coincident spikes of E-I pairs before and after shuffling of trials. Single I-units and pairs of I-units performed better than single E-units and pairs of E-units, respectively (p < 0.05 andp < 0.01, respectively, two-tailedt test). Pairs of cells of the same type performed better than bursts of spikes of single pyramidal cells (p < 0.01 for both E- and I-type neurons). Bursts, in turn, performed better than isolated spikes fired by the respective units (p < 0.01 for both E- and I-type neurons). Feature extraction by opposite-type pairs was close to chance performance (p_E = 0.5).p_E computed for shuffled spike trains was not significantly different from p_Ecalculated for simultaneously recorded spike trains. The mean values ofp_E were computed from the lowest values ofp_E observed regardless of the size of the best time window. Time windows of <5 msec were not used, because the number of spikes coinciding within such a time frame was too small to yield reliable results (Fig. 4c). Error bars indicate SEM. The numbers above the bars give the overall number of stimulus conditions (cutoff frequencies and contrasts) for all cells or cell pairs analyzed.pairs-nsh, Simultaneously recorded spike trains (trials not shuffled); pairs-sh, pair data with trials shuffled;single-bursts, burst spikes of single pyramidal cells;single-isol., isolated spikes of single cells.

Fig. 7.

Terminal spread of P-receptor afferents.Top, Transverse sections at hindbrain level in two preparations (left, right, respectively). The locations of the terminal fields of two Neurobiotin-filled P-receptor afferent fibers within CM are indicated by the boxes.Bottom, Magnified views of the respective cells. In both cases, the terminal fields were reconstructed from three consecutive transverse sections (thickness, 50 μm) of the ELL. The section at theleft corresponds to level −6, and the section at theright corresponds to level −9 of the brain atlas ofMaler et al. (1991). C, Cerebellomedullary cistern;CCb, corpus cerebelli; CM, centromedial segment of ELL; CL, centrolateral segment of ELL;d, dorsal; g, granular cell layer of ELL;l, lateral; L, lateral segment of ELL;M, medial segment of ELL; MLF, medial longitudinal fasciculus.