[Paper Review] On the role of population heterogeneity in emergent communication

Mathieu Rita, Florian Strub+3 more|arXiv (Cornell University)|Apr 27, 2022

Language and cultural evolution4 citations

TL;DR

This paper investigates why larger populations in neural emergent communication simulations fail to produce more structured languages, despite sociolinguistic theory predicting otherwise. By introducing heterogeneous training speeds among agents—particularly speaker-listener asymmetry—it demonstrates that population heterogeneity, not size alone, drives the emergence of more stable and structured communication protocols, resolving a long-standing contradiction in the literature.

ABSTRACT

Populations have often been perceived as a structuring component for language to emerge and evolve: the larger the population, the more structured the language. While this observation is widespread in the sociolinguistic literature, it has not been consistently reproduced in computer simulations with neural agents. In this paper, we thus aim to clarify this apparent contradiction. We explore emergent language properties by varying agent population size in the speaker-listener Lewis Game. After reproducing the experimental difference, we challenge the simulation assumption that the agent community is homogeneous. We first investigate how speaker-listener asymmetry alters language structure to examine two potential diversity factors: training speed and network capacity. We find out that emergent language properties are only altered by the relative difference of learning speeds between speaker and listener, and not by their absolute values. From then, we leverage this observation to control population heterogeneity without introducing confounding factors. We finally show that introducing such training speed heterogeneities naturally sort out the initial contradiction: larger simulated communities start developing more stable and structured languages.

Motivation & Objective

• To investigate why larger simulated populations in emergent communication fail to produce more structured languages, despite sociolinguistic evidence suggesting otherwise.
• To challenge the assumption of homogeneous agent populations in existing neural simulation frameworks.
• To examine how agent-level asymmetries—particularly relative training speeds—impact language structure and stability.
• To demonstrate that introducing controlled heterogeneity in training dynamics can restore the expected positive effect of population size on language quality.
• To provide a computational explanation for the discrepancy between sociolinguistic observations and recent neural emergent communication results.

Proposed method

• The authors use the Lewis referential game as the core environment, where speakers describe objects to listeners to achieve mutual understanding.
• They vary population size (N=2 to N=10) while maintaining identical training hyperparameters in homogeneous settings to isolate the effect of size.
• They introduce controlled heterogeneity by assigning different learning speeds to speakers and listeners, modeling real-world asymmetries in learning capacity.
• Heterogeneity is systematically varied using a parameter σp that controls the standard deviation of training speeds across agents.
• Language quality is evaluated using metrics including success rate, compositionality, and generalization, with results averaged across multiple random seeds.
• The authors compare homogeneous vs. heterogeneous populations across different population sizes to isolate the impact of diversity on language emergence.

Experimental results

Research questions

• RQ1 Why do larger populations in neural emergent communication simulations fail to produce more structured languages, despite sociolinguistic evidence to the contrary?
• RQ2 To what extent does agent-level heterogeneity—specifically differences in training speed—mediate the relationship between population size and language structure?
• RQ3 Can introducing controlled training speed diversity in neural agents restore the expected positive correlation between population size and language quality?
• RQ4 How does relative training speed between speaker and listener roles influence the stability and compositionality of emergent communication protocols?
• RQ5 Does population heterogeneity reduce variance in language outcomes, leading to more reliable and systematic communication?

Key findings

• In homogeneous populations, increasing population size from N=2 to N=10 leads to a decline in average language scores, indicating no benefit from larger groups.
• The relative training speed between speaker and listener is a critical factor: absolute speeds matter less than their relative difference.
• When training speed heterogeneity is introduced (σp > 0), language scores improve significantly with increasing population size, with gains in both mean performance and reduced variance.
• For N=10, heterogeneous populations achieve language quality close to the best-performing N=2 pair, suggesting convergence toward optimal communication protocols.
• The improvement is most pronounced in metrics like compositionality and generalization, indicating more systematic and robust language emergence.
• The results show that population heterogeneity—specifically in learning dynamics—acts as a key mechanism enabling larger communities to develop structured languages, resolving a long-standing empirical contradiction.