[Paper Review] Focus on What's Informative and Ignore What's not: Communication Strategies in a Referential Game

Roberto Dessì, Diane Bouchacourt+2 more|arXiv (Cornell University)|Nov 5, 2019

Language and cultural evolution4 citations

TL;DR

This paper investigates how neural agents develop communication strategies in a referential game by analyzing their object-symbol mappings under different environmental distributions. Using mutual information to quantify feature relevance, the authors show that agents ignore less informative, non-uniformly distributed features and instead optimize communication by focusing on more informative, uniformly distributed features—resulting in a 26–30% increase in mutual information for remaining features, suggesting that non-uniformity can guide more efficient language emergence.

ABSTRACT

Research in multi-agent cooperation has shown that artificial agents are able to learn to play a simple referential game while developing a shared lexicon. This lexicon is not easy to analyze, as it does not show many properties of a natural language. In a simple referential game with two neural network-based agents, we analyze the object-symbol mapping trying to understand what kind of strategy was used to develop the emergent language. We see that, when the environment is uniformly distributed, the agents rely on a random subset of features to describe the objects. When we modify the objects making one feature non-uniformly distributed,the agents realize it is less informative and start to ignore it, and, surprisingly, they make a better use of the remaining features. This interesting result suggests that more natural, less uniformly distributed environments might aid in spurring the emergence of better-behaved languages.

Motivation & Objective

• To understand how artificial agents develop communication strategies in a referential game based on environmental feature distributions.
• To investigate whether agents prioritize salient, informative features over less informative ones when forming emergent language.
• To quantify the relationship between object feature distributions and the resulting message protocol using information-theoretic measures.
• To explore whether non-uniform environmental distributions can lead to more efficient and structured communication protocols.

Proposed method

• Two neural network-based agents (sender and receiver) play a referential game using 5-dimensional discrete object vectors.
• The sender encodes the target object and generates a message via a Gumbel-Softmax-regularized RNN, while the receiver decodes the message to identify the target from a distractor.
• The agents are trained using cross-entropy loss with Adam optimization over 50 epochs, with vocabulary size fixed at 1100 symbols.
• Mutual information (MI) is computed between each object feature and the message distribution to assess feature utilization.
• Two environments are compared: one with uniformly distributed features (each value 1–4 with 0.25 probability), and another with a skewed first feature (value 1 with 0.75 probability).
• Performance is evaluated via accuracy, message entropy, and unique message/target vector counts to assess protocol complexity and generalization.

Experimental results

Research questions

• RQ1 Do agents selectively focus on informative features when describing objects in a referential game?
• RQ2 How does the distribution of features in the environment affect the agents’ choice of features for communication?
• RQ3 Does making a feature less informative (non-uniform) lead agents to ignore it and reallocate communication resources to other features?
• RQ4 To what extent does non-uniformity in feature distribution improve the efficiency of the emergent communication protocol?
• RQ5 Can mutual information between object features and message distributions serve as a reliable metric for evaluating feature utilization in emergent language?

Key findings

• In the uniformly distributed environment, agents achieve 98.61% accuracy but rely on an arbitrary subset of three features, as evidenced by similar mutual information (MI) values across features (mean 0.61 ± 0.41).
• When one feature is made non-uniform (75% probability for value 1), agents reduce its MI to 0.07 ± 0.06, indicating effective ignoring of the uninformative feature.
• The remaining four features (2–5) show a 26–30% increase in average MI compared to the uniform case, suggesting agents make more effective use of the remaining features.
• Despite fewer unique target vectors (772 vs. 1005) in the skewed environment, the number of unique messages is only 8% lower (46 vs. 50), indicating higher message efficiency.
• The entropy of the message distribution is slightly lower in the skewed environment (5.35 vs. 5.46), indicating a more focused and less random protocol.
• The agents maintain comparable performance (98.52% accuracy in skewed vs. 98.61% in uniform), showing that ignoring uninformative features does not degrade task performance.