Associative learning enhances population coding by inverting interneuronal correlation patterns

Abstract

Learning-dependent cortical encoding has been well described in single neurons. But behaviorally relevant sensory signals drive the coordinated activity of millions of cortical neurons; whether learning produces stimulus-specific changes in population codes is unknown. Because the pattern of firing rate correlations between neurons--an emergent property of neural populations--can significantly impact encoding fidelity, we hypothesize that it is a target for learning. Using an associative learning procedure, we manipulated the behavioral relevance of natural acoustic signals and examined the evoked spiking activity in auditory cortical neurons in songbirds. We show that learning produces stimulus-specific changes in the pattern of interneuronal correlations that enhance the ability of neural populations to recognize signals relevant for behavior. This learning-dependent enhancement increases with population size. The results identify the pattern of interneuronal correlation in neural populations as a target of learning that can selectively enhance the representations of specific sensory signals.

Figure 1

Experimental design. (A) Schematic of behavioral training apparatus. (B) Stimulus design. Each letter denotes a single motif. Task-relevant motifs (green) indicated whether to respond left or right. Task-irrelevant motifs (red) were paired in sequence with task-relevant motifs and occurred with equal probability in left stimuli and right stimuli. Novel motifs (black) were never presented during behavioral training, but were presented during neural recording. (C) Example “go left” stimulus (top) and “go right” stimulus (bottom). Arrowheads denote 20ms silent periods between motifs. (D) Mean (±SEM) acquisition curve showing increase in performance with training. Blocks are non-overlapping. Dots at right denote behavioral performance for each bird during the 200 trials immediately prior to neural recording. (E) Summary of responses to probe trials. Each probe trial was a single motif presented in isolation (i.e. not paired) and not reinforced. Data shown here are the percentages of all trials that elicited left responses. Each point represents a single bird. Task-relevant motifs in isolation show a strong learned association with left and right responses. Task-irrelevant motifs in isolation show no learned association, with each motif eliciting similar responses, on average. Although motifs were counter-balanced between birds, the labels for task-relevant and task-irrelevant motifs are given as A–D and E–H, respectively, for ease of display and interpretation. The mean task-irrelevant responses were intermediate to the go left and go right responses for all subjects. (F) As in (E) but for the total number of responses (both left and right). Task-irrelevant motifs elicited response rates nearly as high as task-relevant motifs. Birds thus associate each task-relevant motif with a pecking in a specific port, but associate each task-irrelevant motif only with the general response of pecking. (G) Schematic of CLM within the avian auditory forebrain circuitry. Arrows denote major projection patterns. Hp: hyperpallium; CMM: caudomedial mesopallium. See also Figure S2.

Figure 2

Example responses from pairs of simultaneously recorded CLM neurons. (A,B) Spiking activity in response to task-relevant motifs (A) and task-irrelevant motifs (B). The top half of each panel shows spike rasters for repeated trials. Black dots denote spikes from neuron 1 and red dots denote spikes from neuron 2. The bottom half of each panel shows PSTHs for each neuron superimposed over the spectrogram of the motif stimulus. (C,D) Mean ± S.D. (averaged over the repeated trials) responses of the same neuron pair from panels A and B to the four task-relevant motifs (C) and to the four task-irrelevant motifs (D). Error bars denote standard deviations. (E,F) Normalized responses to each of motif showing no noise correlations for task-relevant motifs (E) and positive noise correlations for task-irrelevant motifs (F). (G–-J) Same as (C–F) but for a second example pair of CLM neurons.

Figure 3

Task-relevance does not influence average noise or signal correlations. (A) Distributions of noise correlations for task-relevant motifs (green), task-irrelevant motifs (red), and novel motifs (black). (B) Same as (A) but for signal correlations. (C) Scatterplot comparing noise correlations for task-relevant motifs with noise correlations for task-irrelevant motifs. Each point denotes one pair. See also Figure S3.

Figure 4

Task-relevance alters the relationship between signal and noise correlations. (A–C) Scatter plots of signal and noise correlations for all pairs for responses to each class of motifs. Each point denotes one neuron pair. Black line depicts zero noise correlation; heavy colored lines are linear regression lines. (D) Mean (±SEM) noise correlations for all pairs with signal correlation greater than 0.4 for relevant (n = 60 pairs), irrelevant (n = 72 pairs), and novel (n =59 pairs) motifs. (E) Mean (±SEM) noise correlations for all pairs with signal correlation less than −0.4 for relevant (n = 53 pairs), irrelevant (n = 39 pairs), and novel (n = 51 pairs) motifs. Wilxocon rank-sum test * p < 0.05, ** p < 0.002.

Figure 5

Motif classification using multinomial logistic regression. Top row shows the two-dimensional response distributions of a pair of neurons for each of the four task-relevant motifs. Probability of observing each pair of responses is represented in grayscale. Bottom row depicts the optimal decoder model fit to the data. The probability of classifying a pair of responses from these two neurons as each of the four motifs is represented in grayscale. For this pair of neurons, the average probability of correct motif classification was 0.507.

Figure 6

Relationship between motif decoding and inter-neuronal correlations. (A) The relationship between signal and noise correlations predicts the effect that correlations have on motif decoding. Neuron pairs are binned according to their signal and noise correlations; the relative performance of the decoder (ratio of correct classifications to correct classifications with trials shuffled) is plotted on the ordinate. The mean classification ratio is greater than 1 for pairs with negative signal and positive noise correlations and for pairs with positive signal and negative noise correlations. In contrast, the ratio is less than 1 for pairs with negative signal and negative noise correlations and for pairs with positive signal and positive noise correlations. T-tests: * p < 0.05; ** p < 1e-9. (B) On average, the classification ratio for task-relevant motifs is greater than task-irrelevant motifs and novel motifs. T-tests: ** p < 0.01. Data are reported as mean ±SEM. See also Figure S1.

Figure 7

Effects of correlations on decoding of task-relevant motifs increase with population size. (A) Mean classification performance for populations ranging in size from 2 to 8 neurons. Task-relevant, task-irrelevant, and novel motifs are denoted by green, red, and black, respectively. Solid lines show performance with correlations intact; dashed lines show performance without correlations (trials shuffled). Although increasing the population size generally increases classification performance even without correlations (dashed lines), the effects of correlations have increasingly larger influences on the task-relevant motifs. (B) Mean (±SEM) classification ratio for each class of motifs as a function of population size. Only the task-relevant motifs exhibit a positive relationship between population size and classification ratio. See also Figure S4.