💬 NLP

SimSiam Naming Game: A Unified Approach for Representation Learning and Emergent Communication

This paper proposes the SimSiam Naming Game (SSNG), a feedback-free framework that replaces inefficient sampling-based updates with symmetric self-supervised representation alignment and Gumbel-Softmax relaxation, enabling agents to learn emergent communication that achieves superior classification performance on complex visual datasets compared to existing methods.

Original authors: Nguyen Le Hoang, Tadahiro Taniguchi, Fang Tianwei, Akira Taniguchi

Published 2026-04-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nguyen Le Hoang, Tadahiro Taniguchi, Fang Tianwei, Akira Taniguchi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine two aliens, Alex and Ben, who have just landed on Earth. They don't speak English, they don't have a dictionary, and they've never met before. Their only goal? To figure out what they are looking at and agree on a name for it.

This is the core problem of Emergent Communication: How do two intelligent beings create a language from scratch just by looking at the same things?

The paper introduces a new method called SimSiam Naming Game (SSNG) to solve this. Here is how it works, explained through simple analogies.

The Old Way: The "Guessing Game" (MHNG)

Before this new method, scientists used a system called the Metropolis-Hastings Naming Game (MHNG).

The Analogy: Imagine Alex and Ben playing a game where Alex guesses a word for a picture. Ben listens and says, "That's close, but maybe try a different word." If Ben accepts the word, they move on. If not, Alex has to guess again.
The Problem: In the real world (with complex images like cats, cars, or trees), there are billions of possible words. This "guess and check" method is like trying to find a specific grain of sand on a beach by picking up one grain at a time and asking, "Is this the one?" It takes forever, and most guesses get rejected. It's too slow and inefficient for complex visual data.

The New Way: The "Mirror Dance" (SSNG)

The authors propose SSNG, which is inspired by a technique called SimSiam (a method used in AI to learn without teachers).

Instead of guessing and checking, Alex and Ben perform a Mirror Dance.

The Setup: Alex and Ben look at the same object (say, a dog) but from slightly different angles or with different filters (like one sees it in black and white, the other in color).
The Secret Code: Each agent has a "brain" that turns the picture into a secret internal code (a latent representation).
The Exchange: They don't just say "Dog." They generate a short, discrete code (like a string of numbers or symbols) based on their internal understanding.
- Alex sends his code to Ben.
- Ben sends his code to Alex.
The Alignment (The Magic): Here is the clever part.
- Alex tries to make his own internal code match the code he received from Ben.
- Ben tries to make his own internal code match the code he received from Alex.
- Crucially: They do this without a teacher telling them "Right" or "Wrong." They just try to make their internal worlds align with each other's messages.

Think of it like two people trying to tune their radios to the same frequency. They don't need a station manager to tell them the right number; they just keep adjusting their dials until the static clears and they hear the same music.

Why is this better?

No Rejection: In the old "Guessing Game," most guesses were wrong and wasted time. In the "Mirror Dance," every interaction helps them get closer to the truth. It's like sliding down a smooth slide instead of climbing a mountain.
Discrete Symbols: Even though the math is continuous (smooth), the agents learn to use distinct "symbols" (like words) rather than just vague feelings. They learn to say "Bark" instead of "Woof-woof-ish."
Heterogeneity: It works even if Alex and Ben have different "brains" (different camera types or neural networks). They still manage to agree on a language.

The Results: Did they learn?

The researchers tested this on famous image datasets (CIFAR-10 and ImageNet).

The Test: They took the "words" the agents invented and asked a simple computer: "Can you guess what the picture is just by reading these words?"
The Winner: The agents using SSNG created a language that was much better at describing the pictures than the old methods. Their "words" were so clear that a simple computer could identify the objects with high accuracy.
The Proof: When they visualized the "words" on a map, the words for "cats" clustered together, and "dogs" clustered together. The agents had successfully invented a structured, meaningful language.

The "Double Identity" Twist

The paper has a cool twist. The authors showed that if you take these two agents and glue them together into one giant system, the whole thing behaves exactly like SimSiam, a famous AI model for learning without labels.

What this means: SSNG isn't just a communication game; it's a new way of teaching AI to understand the world. It proves that communication is a powerful tool for learning. By trying to understand each other, the agents actually learn to understand the world better.

Summary

SimSiam Naming Game is a new way for AI agents to learn a language without a teacher. Instead of playing a slow, frustrating guessing game, they perform a "Mirror Dance," constantly adjusting their internal thoughts to match their partner's messages. This process is fast, efficient, and results in a shared language that is surprisingly good at describing complex visual concepts. It shows that learning to talk to each other is one of the best ways to learn about the world.

1. Problem Statement

Emergent Communication (EmCom) investigates how autonomous agents develop symbolic languages through interaction without predefined rules. While frameworks like the Metropolis–Hastings Naming Game (MHNG) have successfully modeled feedback-free communication via joint attention, they suffer from significant limitations:

Sample Inefficiency: MHNG relies on sampling-based updates (Metropolis–Hastings) to approximate Bayesian inference. In high-dimensional perceptual spaces (e.g., complex visual datasets), this leads to excessively high rejection rates, making learning slow and difficult to scale.
Feedback Dependency: Many alternative approaches (Referential Games, Reconstruction Games) rely on explicit task-based feedback (rewards, listener accuracy, or reconstruction loss), which contradicts the developmental psychology insight that early language acquisition often occurs through interaction cues (like joint attention) rather than explicit correction.

The core challenge is to design a feedback-free EmCom framework that is sample-efficient, capable of handling high-dimensional visual data, and capable of learning discrete symbolic representations that capture semantic structure.

2. Methodology: SimSiam Naming Game (SSNG)

The authors propose SSNG, a framework that unifies EmCom with Self-Supervised Learning (SSL), specifically drawing inspiration from the SimSiam architecture.

Core Architecture

Agent Structure: Two autonomous agents (A and B) are implemented as symmetric branches of a Siamese network. Each agent consists of:
- Perception Encoder ( $f$ ): Encodes an observed input ( $x$ ) into a latent representation ( $z$ ).
- Language Coder: A bidirectional module containing a Generator ( $h$ ) (latent $\to$ message) and an Interpreter ( $l$ ) (message $\to$ latent).
Discrete Communication: To enable end-to-end gradient-based optimization while maintaining discrete symbols, messages are generated using Gumbel–Softmax relaxation with a Straight-Through estimator.
Interaction Loop:
1. Agents observe different views of the same object ( $x_A, x_B$ ).
2. They encode these into latent states ( $z_A, z_B$ ).
3. They generate discrete messages ( $w_A, w_B$ ) and exchange them.
4. Each agent interprets the partner's message to reconstruct a latent state.
5. Agents update their parameters to align their internal representations with the inferred representations from the partner's message.

Learning Objective

The framework is derived from a Variational Inference (VI) perspective, treating symbol emergence as the alignment of latent representations mediated by message exchange. The objective function ( $J_{SSNG}$ ) maximizes an Evidence Lower Bound (ELBO) and consists of three key terms (derived from the Collective Predictive Coding hypothesis):

Collective Regularization: Encourages consistency between an agent's emitted message and the message received from the partner (promoting a shared lexicon).
Individual Prediction Error: Enforces consistency between an agent's internal latent state and the latent state reconstructed from its own generated message.
Individual Regularization: Enforces consistency between an agent's internal latent state and the latent state reconstructed from the partner's message.

Key Mechanism: SSNG utilizes the Stop-Gradient mechanism (central to SimSiam). When Agent A processes Agent B's message, the message tensor is detached from Agent B's computational graph. This prevents representational collapse (where agents output trivial constant messages) and ensures stable, asymmetric gradient flow without requiring negative samples or explicit rewards.

Theoretical Unification

The paper demonstrates that SSNG is a generalization of SimSiam. If the two agents are merged into a single system processing two augmented views of the same image, the SSNG framework mathematically reduces to the standard SimSiam objective. This proves that SSNG preserves the powerful representation learning properties of SSL while extending them to a decentralized, multi-agent communication scenario.

3. Key Contributions

Proposal of SSNG: A novel, feedback-free EmCom framework that replaces inefficient sampling-based updates (MHNG) with a symmetric, self-supervised representation alignment objective.
Theoretical Bridge: Establishes a principled connection between Emergent Communication and Self-Supervised Learning, showing that symbol emergence can be viewed as a latent alignment process similar to SSL.
Discrete Differentiability: Successfully implements discrete symbolic communication within a gradient-based optimization framework using Gumbel–Softmax, preserving the discrete nature of language while allowing for efficient training.
Unified Framework: Demonstrates that merging the two-agent SSNG system recovers standard SimSiam performance, validating the framework's robustness and its ability to extend centralized SSL to decentralized multi-agent settings.

4. Experimental Results

The authors evaluated SSNG on CIFAR-10 and ImageNet-100 (a 20-class subset), comparing it against Referential Games, Reconstruction Games, MHNG, and No-Communication baselines.

Experiment 1: Emergent Communication (EmCom)

Metric: Linear-probe classification accuracy on the discrete messages generated by agents (frozen after training).
Results:
- CIFAR-10: SSNG achieved ~51% accuracy, significantly outperforming MHNG (~~31%), Reconstruction Games (~~36%), and Referential Games (~23%).
- ImageNet-100: SSNG achieved ~43-45% accuracy, again substantially higher than MHNG (~16-22%) and other baselines.
- Alignment: SSNG showed high cross-agent message agreement (Cosine Similarity ~0.82, Overlap ~0.83), whereas MHNG and NoCom showed very low alignment.
Robustness: Experiments with heterogeneous agents (different backbone architectures, e.g., ResNet-18 vs. ResNet-34) showed that SSNG still learned a consistent shared language, proving that alignment is driven by interaction rather than architectural symmetry.

Experiment 2: Representation Learning (SSL)

Setup: The two-agent system was collapsed into a single network processing augmented views (SimSiam mode).
Results:
- On CIFAR-10, SSNG achieved 86.10% Top-1 accuracy, outperforming SimSiam (85.26%), VI-SimSiam, and SimCLR.
- On ImageNet-100, SSNG achieved 78.61%, slightly below SimSiam (79.53%) but outperforming SimCLR and VI-SimSiam.
- Ablation: Removing the stop-gradient mechanism caused performance to collapse to random chance (~10%), confirming that the mechanism is critical for preventing trivial solutions.

5. Significance and Implications

Efficiency: SSNG solves the scalability issue of MHNG in high-dimensional visual domains by eliminating sampling-based rejection, making it feasible to learn emergent languages on complex datasets like ImageNet.
Feedback-Free Learning: It validates the hypothesis that shared symbolic systems can emerge purely through representation alignment and joint attention, without the need for external rewards or explicit error signals.
Unified Theory: The work provides a mathematical unification of language games and self-supervised learning, suggesting that the mechanisms driving modern SSL (like SimSiam) are fundamentally compatible with, and perhaps essential for, multi-agent communication.
Future Directions: The framework opens new avenues for studying compositional generalization, scaling to larger agent populations, and applying these principles to multimodal and sequential communication tasks.

In conclusion, SSNG represents a significant advancement in EmCom, offering a computationally efficient, theoretically grounded, and empirically superior method for agents to develop shared symbolic languages through interaction alone.