Active sampling as an information seeking strategy in primate vocal interactions

Abstract

Active sensing is a behavioral strategy for exploring the environment. In this study, we show that contact vocal behaviors can be an active sensing mechanism that uses sampling to gain information about the social environment, in particular, the vocal behavior of others. With a focus on the real-time vocal interactions of marmoset monkeys, we contrast active sampling to a vocal accommodation framework in which vocalizations are adjusted simply to maximize responses. We conduct simulations of a vocal accommodation and an active sampling policy and compare them with actual vocal interaction data. Our findings support active sampling as the best model for real-time marmoset monkey vocal exchanges. In some cases, the active sampling model was even able to partially predict the distribution of vocal durations for individuals to approximate the optimal call duration. These results suggest a non-traditional function for primate vocal interactions in which they are used by animals to seek information about their social environments.

Fig. 1. Vocal behavior within a single dyadic interaction is diverse and dynamic.

A Experimental set-up of the occluded and dyadic vocal interactions. B–E Vocal properties for individuals A and B, respectively. B Scatter plot for each individual with the time that each call was produced in the x-axis and the call duration of said call in the y-axis. The call duration is calculated as the total duration of a sequence of vocal syllables (continuous bouts of vocalization) with less than 1 s between each other, as exemplified in the spectrogram. C Value of Bayesian Information Criterion (BIC) versus the number of components for each individual when clustering with a Gaussian mixture model the duration of all vocalizations produced after ¼ of the session. For both individuals, 3 is the optimal number of clusters. D Probability density function of each of the clusters for each individual, as determined by the Gaussian mixture model of the optimal number of clusters. E Dynamics of the time-binned coefficient of variation (CV, standard deviation divided by the mean) of the call durations for each individual, calculated via splitting the data into time windows, and calculating the CV in each window. The solid line represents a sigmoid fit of the obtained CVs. F Scatter plot of vocal durations for the population of six individuals throughout the sessions. G BIC versus the number of clusters shows that 3 is also the optimal number of clusters of call durations considering the population as a whole. H Probability density function illustrating what are the clusters. I Dynamics of the CV of call duration for the population. A solid line is a sigmoid fit.

Fig. 2. Different vocalizations chosen by a policy lead to different degrees of information acquisition.

In both models, the agent starts with a belief (represented by a distribution) of what call duration leads to the highest response probability. Using a policy applied to this belief (i.e. the rules determining what action to take), the agent emits a vocalization with a specific duration. Based on the information it gets from hearing or not a response, it updates the belief using the Bayes rule. A In the vocal accommodation policy, the vocalization is chosen to maximize the immediate response probability (i.e., the duration corresponding to the peak of the belief). B Another possibility for the policy is the active sampling policy, in which the vocalization is chosen to maximize the learning (see methods). In the example in the figure, the policy leads to vocalizations on the highest slopes of the belief, rather than the peak, because they have a higher learning potential, achieving an update in the belief with higher magnitude and thus a narrower updated belief than the vocal accommodation model.

Fig. 3. The active sampling model predicts diversity (3 clusters) and dynamics (sudden transition) of vocal interaction.

A Scatter plot of the simulated call durations in an interaction between two agents using a vocal accommodation policy. B Value of Bayesian information criterion (BIC) versus the number of components when Gaussian mixed models are used to cluster the call durations after ¼ of the sessions simulated by the vocal accommodation policy. The optimal number of clusters defined via the elbow method is 2. C Probability density function of each of the clusters from the vocal accommodation model, using the optimal number of clusters. D Dynamics of the time-binned coefficient of variation (CV, standard deviation divided by the mean) of the vocal accommodation simulation, calculated via splitting the data into bins in time, and calculating the CV in each bin. The solid line represents a sigmoid fit of the CVs. E Scatterplot of the simulated call durations in an interaction between two agents using the active sampling policy. F BIC vs number of components when clustering vocal durations with a Gaussian mixture model from calls taken after ¼ of the sessions in active sampling simulation. G Probability density function of each of the clusters from the active sampling model, using the optimal number of clusters given by the elbow method. H Dynamics of the time-binned CV of the active sampling simulation, along with a sigmoid fit.

Fig. 4. Marmoset behavior matches predictions from the active sampling model.

A Graph showing increase in proportion of calls that get a response as the session progresses for the population of 6 individuals, along with 95% CI. B Probability density function of each of the clusters for individual A, as determined by the Gaussian mixture model of the optimal number of clusters shown in Fig. 1D. C Proportion of calls from individual A that got a response (in the x-axis) given different call durations in the y-axis. The dotted line shows alignment of the vocal clusters similarly to the curve, as predicted by the active sampling model in Fig. 2B. D Probability density function of each of the clusters for individual B, as determined by the Gaussian mixture model of the optimal number of clusters shown in Fig. 1D. E Proportion of calls from individual B that got a response (in the x-axis) given different call durations in the y-axis. Dotted line illustrates alignment of the vocal clusters with the curve. F Comparison between call duration with highest chance of response (brown spiky dot) and vocalization clusters for each marmoset (green dot for the clusters around peak response, black for the extra cluster).