Abstract
There is growing interest in the language developed by agents interacting in emergent-communication settings. Earlier studies have focused on the agents' symbol usage, rather than on their representation of visual input. In this paper, we consider the referential games of Lazaridou et al. (2017) and investigate the representations the agents develop during their evolving interaction. We find that the agents establish successful communication by inducing visual representations that almost perfectly align with each other, but, surprisingly, do not capture the conceptual properties of the objects depicted in the input images. We conclude that, if we are interested in developing language-like communication systems, we must pay more attention to the visual semantics agents associate to the symbols they use.
Introduction
In emergent communication research, agents are often tasked with developing a shared language to solve visual referential games. The underlying assumption is that grounding these games in real-world images will lead to symbols that correspond to human-interpretable concepts. However, the internal visual representations that support this communication are rarely scrutinized. This study investigates whether agents' internal "brain" states actually align with the conceptual structure of the world or if they drift toward inscrutable, task-specific patterns.
Methodology
The study utilizes a referential game involving a Sender and a Receiver. The Sender observes a target image and a distractor, then transmits a discrete symbol to the Receiver, who must identify the target. Two game variants were tested:
- Same-image game: Both agents see identical pixel-level images.
- Different-image game: Agents see different instances of the same concept (e.g., two different photos of a dog), theoretically forcing conceptual abstraction.
To evaluate internal states, the authors employed Representational Similarity Analysis (RSA). RSA measures how well the similarity relationships in one space (e.g., raw visual features) are preserved in another (e.g., the agent's internal embeddings).
Key Insights: The Great Representational Drift
A primary finding is that as agents learn to communicate, their internal representations drift significantly from the original visual hierarchy. While the Sender and Receiver's representations align almost perfectly with each other (RSA $\approx$ 0.98), their alignment with the original input features (RSA $\approx$ 0.33) drops sharply.
This suggests that agents are not learning "concepts" as humans understand them. Instead, they develop a private, mutually consistent representational space that ignores the categorical similarities present in the input data.
Results and the "Noise Test"
The most striking evidence of non-conceptual learning comes from the "Noise Test." Agents trained on natural images were tested on their ability to communicate about random Gaussian noise. Surprisingly, they achieved near-perfect accuracy.
As shown in the table, agents trained on the "same-image" task achieved 95% accuracy on noise. Even those trained on the "different-image" task—intended to promote conceptual thinking—maintained 87% accuracy on noise. This proves that agents exploit low-level, high-dimensional relational properties rather than stable conceptual categories like "dog" or "car."
Conclusion
The success of emergent communication in these games is often deceptive. Agents can achieve high task performance by exploiting inscrutable visual patterns that have no human-interpretable semantic meaning. To develop truly language-like systems, future work must design more demanding environments—perhaps involving more distractors or cross-contextual symbol usage—that force agents to remain faithful to the conceptual structure of the perceptual input.
Summary updated