Translating Neuralese

By Jacob Andreas, Anca Dragan, Daniel Klein. 2017

60 citations recorded by the source corpus.

Several approaches have recently been proposed for learning decentralized deep multiagent policies that coordinate via a differentiable communication channel. While these policies are effective for many tasks, interpretation of their induced communication strategies has remained a challenge. Here we propose to interpret agents’ messages by translating them. Unlike in typical machine translation problems, we have no parallel data to learn from. Instead we develop a translation model based on the insight that agent messages and natural language strings mean the same thing if they induce the same belief about the world in a listener. We present theoretical guarantees and empirical evidence that our approach preserves both the semantics and pragmatics of messages by ensuring that players communicating through a translation layer do not suffer a substantial loss in reward relative to players with a common language.

Subfields: Artificial Intelligence, Natural Language Processing, Reinforcement Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Reinforcement Learning

Citations in the corpus

Scalable agent alignment via reward modeling: a research direction (2018, background): “, 2018), translating neural activations into natural language on tasks also performed by humans (Andreas et al., 2017), and combining different interactive visualization techniques (Olah et al.”
Learning Cooperative Visual Dialog Agents with Deep Reinforcement Learning (2017, background): “The more recent resurgence has focused on deep RL [1, 8, 10, 13, 18, 19, 23, 26].”
Emergence of Grounded Compositional Language in Multi-Agent Populations (2017, background): “While there is recent work on interpreting continuous machine languages (Andreas, Dragan, and Klein 2017), the discrete nature and small size of our symbol vocabulary makes it possible to manually labels to the symbols.”
Executing Instructions in Situated Collaborative Interactions (2019, background): “Collaboration has also been studied for emergent communication (e.g., Andreas et al., 2017; Evtimova et al., 2017).”
Human-Centered Artificial Intelligence and Machine Learning (2019, background): “, 2017), or through natural language processing (Andreas et al., 2017).”
Sometimes We Want Ungrammatical Translations (2021, background): “Translation systems that are contingent on preserving syntax and semantics are used as interpretors to decode the interaction between components of a neural network (Andreas et al., 2017).”
Explanation and Justification in Machine Learning : A Survey Or (2017, background): “[ Andreas et al. , 2017 ] also studied messages passed between agents in systems with learned deep communicating policies.”
Inferring Rewards from Language in Context (2022, background): “Andreas et al. (2017); Sumers et al. (2021) have made the same observation: reference games conﬂate whether the speaker’s objective is to inﬂu-ence beliefs or actions, and modeling the speaker as one or the other produces distinct interpretations of utterances (e.g., speakers that only optimize for…”
Emergent Translation in Multi-Agent Communication (2017, background): “& Titov, 2017) and multi-turn communication with a variable-length conversation (Evtimova et al., 2017), and simple compositionality (Kottur et al., 2017; Mordatch & Abbeel, 2017). Meanwhile, Andreas et al. (2017) proposed a model to interpret continuous message vectors by “translating” them”
Compositional Explanations of Neurons (2020, background): “As our compositional formulas increase in complexity, they begin to resemble approaches to generating natural language explanations of model decisions [2, 8, 16, 17, 31].”
Learning to Interactively Learn and Assist (2019, methodology): “…behavior but in task domains that do not require knowledge transfer between the agents, while others (Foerster et al. 2016; Lazaridou, Peysakhovich, and Baroni 2016; Lowe et al. 2017; Andreas, Dragan, and Klein 2017; Mor-datch and Abbeel 2018) all emerge communication over a communication channel.”
Post-hoc Interpretability for Neural NLP: A Survey (2021, background): “This issue is somewhat unique to the NLP literature and is better treated in other fields [6].”
Explainable Neural Computation via Stack Neural Module Networks (2018, background): “by gradient propagation [33,37,27,40,28], post-hoc natural language explanations [15,2] and network dissection [7].”
Emergent Language in a Multi-Modal, Multi-Step Referential Game (2017, methodology): “In the ﬁrst one, communication is used as an essential tool for sharing information among multiple active agents in a reinforcement learning scenario [1–4].”