[Paper Review] Developmentally motivated emergence of compositional communication via template transfer

Tomasz Korbak, Julian Zubek+3 more|arXiv (Cornell University)|Oct 4, 2019

Language and cultural evolution34 references25 citations

TL;DR

This paper proposes a template transfer approach for emergent compositional communication in multi-agent systems, where a receiver trained on non-compositional signaling games is reused to train a new sender, enabling compositional protocols without architectural inductive biases. The method achieves superior zero-shot generalization, context independence, and topographical similarity (0.85) compared to baselines, demonstrating that compositionality can emerge from learned biases in simpler communication protocols.

ABSTRACT

This paper explores a novel approach to achieving emergent compositional communication in multi-agent systems. We propose a training regime implementing template transfer, the idea of carrying over learned biases across contexts. In our method, a sender-receiver pair is first trained with disentangled loss functions and then the receiver is transferred to train a new sender with a standard loss. Unlike other methods (e.g. the obverter algorithm), our approach does not require imposing inductive biases on the architecture of the agents. We experimentally show the emergence of compositional communication using topographical similarity, zero-shot generalization and context independence as evaluation metrics. The presented approach is connected to an important line of work in semiotics and developmental psycholinguistics: it supports a conjecture that compositional communication is scaffolded on simpler communication protocols.

Motivation & Objective

• To investigate whether compositional communication can emerge through adaptation of pre-existing non-compositional communication protocols.
• To explore whether learned biases from simpler games can scaffold the emergence of compositional communication in more complex environments.
• To evaluate the effectiveness of template transfer as a mechanism for inducing compositionality in multi-agent communication systems.
• To compare template transfer with the obverter algorithm in terms of zero-shot generalization, context independence, and topographical similarity.
• To ground the approach in developmental psycholinguistics and semiotics, aligning with the hypothesis that compositionality builds on iconic and indexical precursors.

Proposed method

• A three-phase training regime: (i) training a visual classifier, (ii) training a sender–receiver pair on a non-compositional object naming game with disentangled loss functions, and (iii) transferring the receiver to train a new sender in a compositional signaling game.
• Template transfer involves reusing the pre-trained receiver policy to guide the training of a new sender using standard cross-entropy loss, preserving learned biases.
• The method relies on decomposing the loss function into disentangled components for color and shape prediction, enabling modular learning without requiring input disentanglement.
• Evaluation metrics include topographical similarity (measuring signal structure), context independence (robustness to input variation), and zero-shot generalization (performance on unseen object combinations).
• The approach is contrasted with the obverter algorithm, which requires symmetric roles and identical architectures, and with random and non-pretrained baselines.
• Experiments are conducted on a Lewis signaling game with five colors and five shapes, using a shared visual encoder and discrete symbol sequences for communication.

Experimental results

Research questions

• RQ1 Can compositional communication emerge through transfer of learned biases from simpler, non-compositional signaling games?
• RQ2 Does template transfer lead to better zero-shot generalization than the obverter algorithm or random baselines?
• RQ3 To what extent does the communication protocol exhibit context independence and topographical similarity under template transfer?
• RQ4 Is the emergence of compositionality possible without imposing architectural inductive biases, such as those required by the obverter algorithm?
• RQ5 Can the developmental progression from iconic/indexical to compositional communication be modeled via transfer learning in multi-agent systems?

Key findings

• Template transfer achieved a topographical similarity score of 0.85, significantly higher than the obverter baseline (0.55) and the non-pretrained baseline (0.30), indicating a more structured and compositional signal.
• The method achieved 74% average test accuracy on unseen object combinations, outperforming the obverter (51%) and the non-pretrained baseline (47%), demonstrating strong zero-shot generalization.
• Context independence was 0.18 for template transfer, compared to 0.12 for the obverter and 0.08 for the baseline, indicating greater robustness to input variations.
• The model trained via template transfer showed 100% training accuracy and 48% test accuracy on both color and shape, indicating effective learning without overfitting to the training distribution.
• The communication protocol learned via template transfer exhibited clear compositional structure: symbols consistently mapped to specific attributes (e.g., symbol '8' for magenta when in first position, '8' for box when in second position), as shown in Table 2(b).
• The results support the hypothesis that compositionality can emerge from simpler, pre-existing communication protocols through transfer learning, without requiring architectural inductive biases.