Multilevel animal societies can emerge from cultural transmission

Abstract

Multilevel societies, containing hierarchically nested social levels, are remarkable social structures whose origins are unclear. The social relationships of sperm whales are organized in a multilevel society with an upper level composed of clans of individuals communicating using similar patterns of clicks (codas). Using agent-based models informed by an 18-year empirical study, we show that clans are unlikely products of stochastic processes (genetic or cultural drift) but likely originate from cultural transmission via biased social learning of codas. Distinct clusters of individuals with similar acoustic repertoires, mirroring the empirical clans, emerge when whales learn preferentially the most common codas (conformism) from behaviourally similar individuals (homophily). Cultural transmission seems key in the partitioning of sperm whales into sympatric clans. These findings suggest that processes similar to those that generate complex human cultures could not only be at play in non-human societies but also create multilevel social structures in the wild.

Figure 1. Empirical multilevel network depicting the three nested levels in the sperm whale society off the Galápagos Islands: individuals within social units within vocal clans.

In the social network, modules of individual whales (coloured small nodes) connected by their social relationships (black lines with thicknesses proportional to the time individuals were identified in the same group) define the social units (letter-labelled large nodes). In the overlapped acoustic network, modules of social units connected by the similarity in acoustic behaviour (grey lines whose thicknesses are proportional to multivariate similarity of coda repertoires) represent the vocal clans (blue: Regular clan, characterized by codas with regularly-spaced clicks; red: Plus-One clan, characterized by codas with extended pause before the final click35). Social relationships and acoustic similarities are replotted results from refs , , respectively.

Figure 2. Schematic representation of the agent-based models.

(a) Coda transmission mechanisms are represented as changes in the coda repertoires vector (squares: coda types; colours: frequency of usage: absent=0, always=100%). Calf agents change repertoires three times (between 0 and 2 year old) under one of following mechanisms. (i) Individual learning: newborn agent a starts with an empty coda vector; half of the elements are randomly selected to receive absolute frequencies of usage from a uniform distribution ∈[0,100]. (ii) Genetic inheritance: newborn agent b starts with an empty coda vector, which is filled with the same coda types and frequencies of its mother B. (iii) Oblique social learning: newborn agent c starts with an empty coda vector; at the age 0 year it randomly samples 62 elements (including zeroed elements) from the coda vector of other adult agents, kin-related or not; at ages 1 and 2 years, the calf repeats the process, replacing a portion of sampled elements. For iv–vi, calves gain an initial repertoire via oblique social learning, then at ages 1 and 2 years, the following effects were included. (iv) Homophily: calf d copies from adult agents of the social unit A, which has the highest coda repertoire similarity with its own social unit. (v) Conformism: calf e preferentially copies the coda types with higher frequencies of usage, here the three codas commonly performed by the adults. (vi) Symbolic marking: calves f and g were born in different social units, which have a specific subset of codas (‘symbol') that all members always perform to mark the identity of the unit (the sequences of red codas). Both calves copy codas from other adults, but also deliberately copy their units' ‘symbols'. (b) Oblique social learning (iii) and the additional effects (iv–iv) occurred at the three social levels. (vii) Social unit: calf agents copy only from agents of their own social unit. (viii) Predefined clans: simulation started with predefined clan labels and calves copy from any agent inside of its predefined clan. (ix) Population: calves copy from any agent in the population. In all scenarios, calves had a low individual learning probability (replacing 1 random coda type by a random frequency) per year.

Figure 3. Coda repertoire similarities and clan partitioning across simulated scenarios.

(a) Agent-based models (ABMs) differed in how a coda was transmitted (individual learning, genetic inheritance and social learning), if there was any transmission biases (homophily, conformism and symbolic marking), and the social level at which the transmission operated (population, social units and predefined clans). Columns represent ABMs and filled cells represent the presence of the model features (transmission mechanisms, biases and social levels) indicated in the rows. Colour code denotes similar transmission mechanisms operating at different social levels. (b) Average coda repertoire similarity of all emergent social units. Whiskers represent s.d. (c) Modularity (Q-values) of the resultant acoustic networks from each ABM. Significantly high modularity values (P<0.001) fall outside of the 95% confidence intervals (whiskers) generated by a theoretical model (1,000 replicates) and indicate the emergence of vocal clans, that is, modules of highly connected social units due to high coda repertoire similarity in the acoustic network. Number of emergent vocal clans is listed on the top of the plot; symbol shapes denote the social level where the transmission mechanisms operated (circle: population; square: social unit; and triangle: predefined clans).

Figure 4. Acoustic networks simulated by the 20 agent-based models (ABMs).

Nodes representing social units are connected by links representing coda repertoire similarity. Colour code indicates similar transmission mechanisms operating at different social levels across the ABMs, which differed in the coda transmission process (individual learning, genetic inheritance, social learning), in the presence, type and combination of transmission biases (columns: homophily, conformism, symbolic marking), and in the social level at which the transmission operated (rows: population, social units, and predefined geographically-segregated clans). ABMs 1 and 2 represent the null agent-based models with no social learning, but individual learning and genetic inheritance of codas, respectively. From ABM 3 to 20, all models contain social learning of codas, with and without biases as indicated. Distinct clans (modules in blue shades) emerged only when codas were transmitted by social learning (SL) biased by conformism (C) and homophily (H) operating in tandem at the population (ABM 15), social unit (ABM 16) and pre-defined clan levels (ABM 17). In all remaining scenarios, the acoustic network resembled the null models (ABMs 1–2), with no distinct clans.