34 - Translating Neuralese, with Jacob Andreas

Agent Communication

Jacob Andreas explores the complexities of decentralized agent communication, where agents use a unique language, termed 'neuralese,' to interact. This language, composed of RNN hidden state vectors, enables agents to coordinate tasks like cleaning or driving without predefined communication protocols 1. Andreas highlights the challenge of translating these neural messages into human-understandable language, emphasizing the need to understand the high-level communication strategies being learned 1.

The problem that we're trying to solve here is given access to one of these kind of pre-trained multi-agent systems that uses communication. Can we take that communication and turn it into something that people can understand?

--- Jacob Andreas

He illustrates this with deep communicating policies (DCPs), where agents play games to develop communication strategies, demonstrating the potential of neuralese in practical applications 2.

   

Translation Methods

The translation of neuralese into natural language involves complex methodologies, as Jacob Andreas explains. Unlike traditional machine translation, this process lacks parallel data, requiring innovative frameworks to map agent messages to human language 3. Andreas describes a belief-oriented translation approach, where the focus is on the belief distribution induced in the listener, rather than direct message translation 3.

We build a model that takes us from that message onto the belief that a sort of optimal listener forms about the state that the speaker might be in upon hearing that message.

--- Jacob Andreas

This method leverages the grounding of messages in real-world states, allowing for effective translation even without direct human-agent communication parallels 4.

   

Game Evaluation

Evaluating the effectiveness of neural translation involves experimental games that test communication strategies. Jacob Andreas discusses three games, including a color-based referring expression game and a two-player driving game, to assess how well neuralese messages can be interpreted 5. The evaluation focuses on whether the correct belief about the speaker's state is induced in the listener, with results showing promising accuracy 5.

You do a lot better, maybe unsurprisingly in this setting, by explicitly doing this kind of belief-oriented view of translation.

--- Jacob Andreas

Additionally, Andreas notes that integrating translation layers between robots and humans can sometimes outperform human-only communication, highlighting the potential of neural translation in practical applications 6.