Multi-Agent Cooperation and the Emergence of (Natural) Language
Reference: Angeliki Lazaridou, Alexander Peysakhovich, and Marco Baroni (2017). ICLR 2017. Source file: 1612.07182v2.pdf. URL
Summary
Introduces a referential-game framework for studying emergent communication: a sender sees a target/distractor image pair and sends a single symbol from a fixed vocabulary; a receiver must identify the target using that symbol. The agents are blank-slate neural networks trained only by communication-success reward. The paper studies whether agents converge, whether the emergent symbols align with human-interpretable semantics, and how to nudge the system toward natural-language-compatible codes.
Two sender architectures (agnostic vs informed) are compared; the informed sender produces richer vocabulary usage. A supplementary supervised image-labelling objective is shown to ground agent symbols to human concepts, making them partially interpretable to crowd-workers.
Key Ideas
- Referential games as minimal test-beds for emergent protocols
- Informed (feature-aware) sender yields more human-like symbol usage
- Symbol purity measured against conceptual (McRae) categories
- Mixing self-play with supervised labelling grounds emergent codes to natural language
- Foundational for later emergent-communication literature
Connections
Conceptual Contribution
- Claim: Two neural-network agents playing a referential game (sender/receiver over image pairs) can develop a symbolic code from scratch; with architectural and supervisory nudges, that code can be made to align with human-interpretable object categories.
- Mechanism: Lewis-style signalling game with REINFORCE training; contrast agnostic vs informed sender architectures; analyse symbol-to-category purity; then mix supervised image-labelling with self-play to ground emergent symbols in human vocabulary (AlphaGo-inspired). Crowdsourced evaluation shows humans can guess the correct image 68% of the time from emitted symbols.
- Concepts introduced/used: Emergent Communication, Referential Games, Lewis Signalling Games, Language Games, Grounding in Human Language, Symbol Grounding Problem, Cheap Talk, Symbol-Category Purity, REINFORCE, Compositionality
- Stance: empirical / deep-learning
- Relates to: Companion/precursor to Emergence of Grounded Compositional Language in Multi-Agent Populations (physical grounding, >2 agents). Feeds the emergent-protocol thesis of A Scalable Communication Protocol for Networks of LLMs. Contrasts with stipulated-semantics ACLs (KQML as an Agent Communication Language, FIPA-ACL).
Tags