Catalytic Role Of Noise And Necessity Of Inductive Biases In The Emergence Of Compositional Communication

Łukasz Kuciński, Tomasz Korbak, Paweł Kołodziej, Piotr Miłoś

TL;DR

This paper demonstrates that for compositional communication to emerge in multi-agent systems, inductive biases on both the data and the training framework are strictly necessary. Crucially, the authors prove and experimentally verify that injecting a moderate amount of noise into the communication channel acts as a powerful catalyst, naturally driving agents to develop structured, compositional languages.

What Problem Does This Paper Address?

In emergent communication studies, researchers train artificial agents to interact and create their own protocols, often using "signaling games" where a Sender describes an object and a Receiver must guess its features. A holy grail in this field is the emergence of compositionality—where complex signals are broken down into simpler, reusable subparts (much like how "red circle" combines the independent concepts of color and shape).

While many believe compositionality can be learned in a purely unsupervised, data-driven manner, this paper introduces a sobering impossibility result: without explicit priors or inductive biases, a unique compositional language cannot be inferred from data alone. The authors tackle the challenge of identifying exactly what mechanisms and biases are required to reliably spark the spontaneous emergence of compositional language.

Methods

The authors utilize a standard signaling game framework but introduce two critical components:

  1. A Noisy Channel: Messages sent from the Sender to the Receiver pass through a channel that randomly scrambles individual symbols with a probability $\epsilon$.
  2. A Factorized Loss Function: The agents are evaluated using a loss function that penalizes mistakes based on the Hamming distance (the number of incorrect features identified), rewarding partially correct guesses.

To train the neural networks end-to-end despite the discrete nature of communication symbols, they employ the Gumbel-Softmax straight-through estimator.

Theoretically, they prove that minimizing this factorized loss over a noisy channel forces the optimal communication protocol to be compositional. Intuitively, a compositional language isolates errors: if one symbol is corrupted by noise, only one feature prediction is ruined, whereas in a non-compositional language, a single corrupted symbol could completely scramble the Receiver's understanding of the entire object.

Datasets

The training and evaluation rely on two synthetic image datasets featuring independent, factorized attributes:

  • Shapes3d: A standard benchmark containing 3D shapes with attributes like floor color, wall color, object color, scale, shape, and orientation. The authors focus primarily on shape and object hue.
  • Obverter Dataset: A dataset consisting of four distinct shapes (box, cylinder, ellipsoid, sphere) rendered in four colors (blue, cyan, gray, green).

To test the boundaries of data priors, the authors also created scrambled variants of these datasets, where images were chopped into tiles and shuffled, or where the labels were randomly permuted.

Key Insights and Ideas

  • The Impossibility of Purely Unsupervised Compositionality: The authors present a discrete analog to recent disentanglement impossibility results. If the data distribution is uniform, any permutation of features looks identical to the model. Therefore, "human-like" compositionality requires human-aligned architectural or data priors.
  • Noise as a Feature, Not a Bug: Instead of hindering communication, channel noise acts as a catalyst. It creates an evolutionary pressure. Because compositional languages degrade gracefully under noise, agents naturally gravitate toward them to maximize their rewards.
  • The Compositionality-Accuracy Trade-off: There is a distinct "sweet spot" for noise. Too little noise fails to incentivize compositionality, while too much noise overwhelms the channel, preventing the agents from learning an accurate communication protocol at all.

Key Results

  • The Noise Sweet Spot: In experiments on the shapes3d dataset, topographic similarity (a standard metric for compositionality, where higher is better) peaked at a noise level of $\epsilon = 0.1$, reaching a score of 0.87, compared to just 0.79 when no noise was present.
  • Architecture Matters: The strongest bias toward compositionality came from using a multi-headed (factorized) output architecture in the Receiver network, which explicitly mirrors the underlying structure of the features.
  • Variable Noise Benefits: Gradually introducing noise (a noise "warmup") or varying it during training led to more stable learning and higher final compositionality scores.

Limitations and Future Work

  • Restrictive Theoretical Assumptions: The mathematical proofs rely on idealized conditions, such as fixed message lengths and alphabet sizes that perfectly match the number of feature values.
  • Synthetic Data Reliance: The methods were only validated on clean, synthetic datasets (shapes3d and obverter). Testing on complex, real-world data remains an open challenge.
  • Simplified Compositionality Model: The paper assumes features are entirely independent. Real-world languages often exhibit hierarchical or non-linear compositionality, which this framework does not yet capture.
  • Supervised Setup: The training pipeline mimics supervised learning; extending this catalytic role of noise to full multi-agent reinforcement learning scenarios is a crucial next step.

Summary updated

Abstract

Communication is compositional if complex signals can be represented as a combination of simpler subparts. In this paper, we theoretically show that inductive biases on both the training framework and the data are needed to develop a compositional communication. Moreover, we prove that compositionality spontaneously arises in the signaling games, where agents communicate over a noisy channel. We experimentally confirm that a range of noise levels, which depends on the model and the data, indeed promotes compositionality. Finally, we provide a comprehensive study of this dependence and report results in terms of recently studied compositionality metrics: topographical similarity, conflict count, and context independence.
BibTeX
@inproceedings{kucinski2021catalytic,
  author = {Łukasz Kuciński and Tomasz Korbak and Paweł Kołodziej and Piotr Miłoś},
  title = {{Catalytic Role Of Noise And Necessity Of Inductive Biases In The Emergence Of Compositional Communication}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2021},
  url = {https://openreview.net/forum?id=FAm_H6ZcuMv},
}

Loading related content...

Figure