INFORMATION-THEORETIC UNDERPINNINGS OF GEN-ERALIZATION AND TRANSLATABILITY IN EMERGENT COMMUNICATION

Abstract

Traditional emergent communication (EC) methods often fail to generalize to novel settings or align with representations of natural language. Here, we show how controlling the Information Bottleneck (IB) tradeoff between complexity and informativeness (a principle thought to guide human languages) helps address both of these problems in EC. Using VQ-VIB, a recent method for training agents while controlling the IB tradeoff, we find that: (1) increasing pressure for informativeness, which encourages agents to develop a shared understanding beyond task-specific needs, leads to better generalization to more challenging tasks and novel inputs; (2) VQ-VIB agents develop an EC space that encodes some semantic similarities and facilitates open-domain communication, similar to word embeddings in natural language; and (3) when translating between English and EC, greater complexity leads to improved performance of teams of simulated English speakers and trained VQ-VIB listeners, but only up to a threshold corresponding to the English complexity. These results indicate the importance of informational constraints for improving self-play performance and human-agent interaction.

1. INTRODUCTION

We wish to develop artificial agents that communicate in grounded settings, via communication that enables high task utility, generalizability to novel settings, and good human-agent cooperation. Emergent communication (EC) methods, wherein agents learn to communicate with each other in an unsupervised manner by maximizing a reward function, take a step towards this vision by producing agents that use grounded communication (Lowe et al., 2017; 2020; Lazaridou & Baroni, 2020) . While numerous EC methods have succeeded in training agents to communicate with each other to solve a particular task, they still fall short of the vision of generalizable and human-interpretable communication. For example, agents trained to discriminate between two types of images will fail to discriminate between sixteen images (Chaabouni et al., 2021b) , and messages often violate human expectations for meanings (Kottur et al., 2017) . In this work, we take steps towards addressing these limitations by building on the informationtheoretic EC approach of Tucker et al. (2022) . This approach connects EC with the Information-Bottleneck (IB, Tishby et al., 1999) framework for semantic systems (Zaslavsky et al., 2018; Zaslavsky, 2020) , via the vector-quantized variational Information Bottleneck (VQ-VIB) neural architecture (Tucker et al., 2022) . Specifically, VQ-VIB agents are trained to optimize a tradeoff between maximizing utility (how well they perform a task), maximizing informativeness (how well a listener can infer a speaker's meaning, independently of any downstream task), and minimizing communicative complexity (roughly the number of bits allocated for communication). While previous EC methods typically focus on task-specific utility maximization (Lowe et al., 2017) , there is broad empirical evidence suggesting that human languages are guided by the IB informativeness-complexity tradeoff (Zaslavsky et al., 2018; 2019; 2021; 2022; Mollica et al., 2021) . Therefore, we hypothesize that taking into account informativeness could improve EC generalizability to novel settings while adjusting complexity could improve the translatability between EC and human languages. Results from our experiments support this hypothesis. First, we show that encouraging informativeness allows EC agents to generalize beyond their training distribution to handle more challenging tasks and out-of-distribution objects, with VQ-VIB achieving the best performance compared to alternative EC methods. Second, we propose a simple method for translating natural language word embeddings (e.g., GloVe, Pennington et al., 2014) into EC signals and use that to simulate humanagent communication in a cooperative object-discrimination task. We find that team performance for English speakers and trained EC agents improves with the communicative complexity of the EC system, but only up to a certain threshold, which corresponds to the complexity of the English object naming system. Together, our findings suggest that training EC agents while controlling the informativeness-complexity tradeoff, in addition to maximizing utility, may simultaneously support improved self-play performance as well as human-agent interaction.

2. RELATED WORK

Our work builds upon prior research in emergent communication (EC), wherein cooperative agents are trained to maximize a reward function. For example, a speaker may observe a "target" image that a listener must identify from a pool of candidate images, based only on the speaker's emitted communication. Researchers in EC often consider the effects of different neural network architectures, training losses, or environmental factors on learned communication (Kottur et al., 2017; Mu & Goodman, 2021; Kuciński et al., 2021; Tucker et al., 2021) . In this work, we are primarily concerned with how information-theoretic properties of agents' communication allow them to generalize or align with human languages, although we also improve upon an existing neural architecture.

2.1. GENERALIZATION OF EMERGENT COMMUNICATION

Several works consider the ability of EC agents to generalize to different settings than those in which they were trained. For example, Lazaridou et al. ( 2018) train agents with symbolic inputs and use novel combinations of such inputs to test generalization. Other researchers use similar train-test gaps to measure the compositionality and generalizability of communication, but such experiments are necessarily restricted to domains with symbolic structure (as those allow recombination of features) (Andreas, 2019; Kuciński et al., 2021; Spilsbury & Ilin, 2022) . Conversely, Chaabouni et al. (2021b) tests agents' abilities to generalize in image-based reference games to 1) more distractor images or 2) inputs from a distinct dataset of images. They find that using harder tasks at training time is important for improving test-time performance and that using population-based voting improves cross-domain transfer. We are similarly interested in the ability of agents to generalize to harder tasks and novel inputs, but we focus on training only a single team of agents; research in learning dynamics due to population effects or changes in environment is complementary to our work.

2.2. LINKS TO NATURAL LANGUAGE

Complementing research of EC in self-play (evaluating agents that were trained together), many researchers explore connections between EC and natural language. Some top-down methods combine pretrained language models with finetuning in grounded environments, but such methods can suffer from "drift" wherein agents learn to ascribe new meanings to words (Lewis et al., 2017; Lazaridou et al., 2020; Jacob et al., 2021) . Other researchers train agents in self-play and then seek to connect learned EC to human-interpretable concepts or natural language (Andreas et al., 2017; Kottur et al., 2017) . Tucker et al. ( 2021) take a middle ground by training agents to communicate via discrete representations in a continuous space, which can be linked to natural word embeddings via supervision during training. We seek to uncover methods for best connecting EC and human communication via translation, with no supervision data when training the agents. Based on information-theoretic analysis of human naming systems, we investigate whether producing EC that matches humans' complexity and informativeness enables better translation.

3. BACKGROUND: INFORMATION-THEORETIC EMERGENT COMMUNICATION

Our work builds on the information-theoretic framework of Zaslavsky et al. (2018) for semantic systems, and especially on its extension to scalable emergent communication in artificial neural agents, proposed by Tucker et al. (2022) . Below we review the theoretical and empirical foundations of this line of work.

