PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Human-Agent Communication via the Information Bottleneck Principle

Mycal Tucker, Julie Shah, Roger Levy, Noga Zaslavsky

arXiv:2207.00088v1cs.AIcs.CL

TL;DR

Emergent communication often prioritizes task utility, while human languages are argued to evolve under pressure to compress meanings into signals through an informativeness–complexity tradeoff [10]. This paper proposes VQ-VIB and trains agents to balance utility, informativeness, and complexity. Across the reported settings, informativeness rewards improve convergence, complexity penalties yield human-like communication, and VQ-VIB performs better than other discrete methods at matched complexity and informativeness.

  • Problem

    Emergent-communication research often optimizes task-specific utility, while human-language accounts emphasize efficient compression through an informativeness–complexity tradeoff.

    Source: §I · p. 1

  • Method

    VQ-VIB trains neural agents to compress inputs into discrete signals embedded in a continuous space while optimizing utility, informativeness, and complexity.

    Source: §I · p. 1, B. Vector Quantized Variational… · p. 2

  • Results

    Across the reported architectures and settings, informativeness rewards improve communication convergence, complexity penalties produce more human-like communication, and VQ-VIB achieves higher utility than other discrete methods at matched complexity and informativeness.

    Source: §I · p. 1, A. Particle Worlds · p. 4, §V · p. 4

  • Takeaways & Limitations

    The findings suggest that cognitive-science-inspired constraints can guide artificial agents toward communication systems with more human-like properties.

    Source: §I · p. 1, §V · p. 4

Abstract

from arXiv · show

Emergent communication research often focuses on optimizing task-specific utility as a driver for communication. However, human languages appear to evolve under pressure to efficiently compress meanings into communication signals by optimizing the Information Bottleneck tradeoff between informativeness and complexity. In this work, we study how trading off these three factors -- utility, informativeness, and complexity -- shapes emergent communication, including compared to human communication. To this end, we propose Vector-Quantized Variational Information Bottleneck (VQ-VIB), a method for training neural agents to compress inputs into discrete signals embedded in a continuous space. We train agents via VQ-VIB and compare their performance to previously proposed neural architectures in grounded environments and in a Lewis reference game. Across all neural architectures and settings, taking into account communicative informativeness benefits communication convergence rates, and penalizing communicative complexity leads to human-like lexicon sizes while maintaining high utility. Additionally, we find that VQ-VIB outperforms other discrete communication methods. This work demonstrates how fundamental principles that are believed to characterize human language evolution may inform emergent communication in artificial agents.

I. INTRODUCTION

Emergent communication often optimizes task utility, but that can yield messages that are difficult for humans to understand or slow to converge. This paper combines utility with informativeness and complexity objectives and introduces VQ-VIB to guide communication toward more human-like systems.

I. INTRODUCTION

  • Task-utility-focused communication may become too complex for humans to understand or converge slowly, motivating additional constraints.The introduction cites prior work for these concerns [7] and [8].Source: §I · p. 1
  • The Information Bottleneck frames human-language efficiency as compressing meanings into signals through a tradeoff between communicative complexity and informativeness [10].Informativeness concerns whether a listener can infer intended meaning independently of a particular task.Source: §I · p. 1
  • VQ-VIB trains agents to trade off task utility, communicative informativeness, and complexity, and the authors report faster convergence with informativeness rewards and human-language-aligned communication under complexity penalties.They also report that VQ-VIB achieves higher utility than other discrete methods at the same complexity and informativeness.Source: §I · p. 1

II. RELATED WORK

Prior emergent-communication work has explored task utility, complexity constraints, and meaningful-message objectives, while color naming studies connect human communication to Information Bottleneck efficiency. This paper explicitly includes utility, informativeness, and complexity in training.

II. RELATED WORK

  • Emergent-communication research generally trains agents on cooperative task rewards, while discretization, noise, or centralized message processing can limit communication complexity [3] [5] [21] [22] [23] [24] [26].Auxiliary losses intended to create meaningful communication have also been associated with improved convergence rates [8] [27] [28].Source: §II · p. 1
  • The framework represents training tradeoffs among utility, informativeness, and complexity with respective weights λU, λI, and λC.The figure caption says these weights guide properties of emergent communication.Source: Figure 1 · p. 2
  • Color naming systems across languages approach Information Bottleneck complexity–informativeness tradeoffs, and AI agents have learned IB-efficient color naming through reference games.Related studies also examine other notions of efficiency.Source: §II · p. 2

III. TECHNICAL APPROACH

The technical approach adapts neural architectures to optimize informativeness and complexity alongside task utility. It builds on an Information Bottleneck framework and uses variational and reconstruction-loss approximations for neural settings.

III. TECHNICAL APPROACH

  • The approach adapts standard neural architectures for informativeness- and complexity-based objectives, introduces VQ-VIB, and combines utility, informativeness, and complexity in an overall objective.Source: §III · p. 2

A. Information Bottleneck for Semantic Communication

  • In the semantic-communication Information Bottleneck framework, speaker and listener jointly trade off message complexity I(X, C) against informativeness measured by expected KL divergence between their mental representations.The resulting principle minimizes I(X, C) − βE[D[m∥m̂]], with β controlling the tradeoff.Source: A. Information Bottleneck for Semantic… · p. 2
  • For neural networks and large-scale domains, the method optimizes a variational upper bound and approximates informativeness with mean squared reconstruction error.The passage compares the upper-bound approach to the ELBO used in VAEs and β-VAEs.Source: A. Information Bottleneck for Semantic… · p. 2

B. Vector Quantized Variational Information Bottleneck

VQ-VIB learns information-constrained discrete communication in a continuous space by combining discrete representation learning with variational information constraints. Its VQ-VAE component maps latent encodings to learned discrete embeddings.

B. Vector Quantized Variational Information Bottleneck

1) Vector-Quantized Variational Autoencoder:

2) β Variational Autoencoder (β-VAE):

The β-VAE encoder maps inputs to Gaussian parameters and samples a latent representation. Its training objective combines reconstruction with a weighted KL penalty.

2) β Variational Autoencoder (β-VAE):

3) VQ-VIB Architecture:

VQ-VIB samples a continuous representation and quantizes it to a discrete signal, while bounding communication complexity. The overall training objective trades off utility, informativeness, and complexity.

3) VQ-VIB Architecture:

  • VQ-VIB samples a continuous representation z from encoder parameters, then maps it to the nearest embedding to produce a discrete communication signal c.The likelihood of c given input x reflects the probability that the sampled representation is closest to that embedding.Source: 3) VQ-VIB Architecture · p. 3

C. Overall Optimization

  • The overall objective weights task utility, informativeness, and low communication complexity, spanning silence, task-optimized signals, and complete descriptions.Source: C. Overall Optimization · p. 3
  • The Figure 5 caption reports that greater λI speeds VQ-VIB convergence in 9 points, while annealing λC lowers reward and complexity across methods.Source: Figure 5 · p. 3
  • In practice, gradient descent optimizes approximations: utility uses policy gradients or supervised loss, informativeness uses negative MSE, and complexity uses a KL bound.The complexity reference distribution is uniform for onehot communication and a unit Gaussian for other methods.Source: C. Overall Optimization · p. 3

A. Particle Worlds

In particle worlds, speakers communicate target locations so listeners can reach landmarks. Higher informativeness improves convergence, while greater complexity penalties reduce both complexity and utility; VQ-VIB outperforms other discrete methods at matched complexity.

A. Particle Worlds

  • Speakers observe target landmarks and communicate to listeners; 9-points uses fixed locations, whereas Uniform spawns landmarks randomly.The team reward is negative distance from the listener to the target.Source: A. Particle Worlds · pp. 3–4
  • VQ-VIB achieves higher utility than other discrete communication methods at the same complexity in both 9-points and Uniform environments.Source: A. Particle Worlds · p. 4
  • Across both environments, higher λI improves convergence, while increasing λC produces less complex, lower-utility communication.Source: A. Particle Worlds · p. 4

B. Color Reference Games

In the color reference game, increasing informativeness improves convergence, and complexity annealing yields a range of near-optimal color-naming systems. VQ-VIB communication vectors also reflect color-space distances, and complexity annealing was necessary to reproduce naming conventions similar to human languages.

B. Color Reference Games

  • Increasing λI at fixed λC improves convergence, while VQ-VIB typically learns more complex communication than onehot.Source: B. Color Reference Games · p. 4
  • As λC increases, complexity decreases, and training points across five seeds closely approximate the IB upper bound for maximum informativeness at each complexity.Source: B. Color Reference Games · p. 4
  • VQ-VIB communication-vector distances correlate with associated CIELAB color distances at r2 = 0.97 ± (0.01), versus r2 = 0.76 ± (0.03) for onehot.Source: B. Color Reference Games · p. 4
  • Annealing complexity was necessary to generate agent naming conventions similar to each of the 110 WCS languages.For equal informativeness at equal complexity, VQ-VIB Before achieved higher utility than onehot.Source: B. Color Reference Games · p. 4

V. CONCLUSION

The conclusion examines utility, informativeness, and complexity together in emergent communication. It reports faster convergence when informativeness is rewarded and more human-like communication when complexity is penalized.

  • Rewarding informativeness led to faster communication convergence, while penalizing complexity created more human-like communication.Source: §V · p. 4
  • The authors describe their work as the first to directly optimize agents for a combination of utility, informativeness, and complexity.Source: §V · p. 4
  • Although the experiments used self-play, the authors identify human-robot communication as a natural extension involving partner-specific communication, complexity constraints, and a small amount of supervised data.Source: §V · pp. 4–5
Loading 2207.00088v1…