Semantic communication: How Orange Is Preparing for the Era of Autonomous Agents

• AI
• Deep learning
• IoT
• Machine learning
• Networks

Share on Twitter : Semantic communication: How Orange Is Preparing for the Era of Autonomous Agents (New window) Share on Linkedin : Semantic communication: How Orange Is Preparing for the Era of Autonomous Agents (New window)

When a human user makes a request—in natural language—to an AI, the AI first calculates its response in a digital representation space and then converts it back to natural language in order to communicate. As a legacy of this language model functionality, communication between autonomous agents—these “small digital brains capable of collaboration”—now mainly takes place via natural language text. The technological choice to require these AI agents, designed to function autonomously to perform certain tasks, to take a detour via English (in most cases) is paradoxical. “This is why, in the future, these exchanges will likely evolve towards direct transmission of semantic representations, enabling more efficient communication and better preserving the semantic richness of information,” explains Louis-Adrien Dufrène, a machine learning researcher at Orange. The idea is to open up new possibilities in terms of functionality and specialization beyond traditional applications.

Semantic communication is therefore the ability to transfer these complex representations between different entities distributed across a network, such as between autonomous agents.

From raw data to meaning: the emergence of vector representations

“Unlike human language, which is composed of words, semantic communications are based on continuous vector representations, also known as embeddings. In concrete terms, this involves transforming information (text, images, sound) into a long list of numbers (a high-dimensional vector). The unique feature of this approach is that it captures the deep structure and precise meaning of the information in a continuous mathematical space. These representations are optimized for specific tasks, and their continuous nature allows them to be finely tuned to application needs. They also make it possible to process different types of data, whether text, images, sounds, or video, in a unified manner. “Semantic communication is therefore the ability to transfer these complex representations between different entities distributed across a network, such as between autonomous agents,” explains Guillaume Larue, a machine learning researcher at Orange.

Networks: transporting high-dimensional complexity

“This development poses several major challenges for telecom operators such as Orange. These semantic representations are high-dimensional vectors (several thousand to tens of thousands) whose direct transmission generates potentially large volumes of data,” explains the researcher. Networks must therefore guarantee low latency and sufficient bandwidth to support these flows. It is therefore necessary to develop specific processing methods for these semantic representations (quantification, compression), or even adapt their very construction, in order to optimize transport without degrading the quality of the information.

Interoperability: towards a universal semantic space?

Different AI models generate distinct semantic representations that are generally incompatible with each other. The lack of standards for exchanging semantic representations therefore creates a risk of market fragmentation, explains Quentin Lampin, machine learning researcher at Orange: “This is why operators such as Orange can play a key role in developing interoperability protocols and semantic translation layers that enable communication between heterogeneous models. ” The question of a universal semantic space, understandable by all models, remains open. While multimodal models are paving the way for shared representations, the challenge is to find the right balance between standardization (for interoperability) and specialization (for performance). “Orange is contributing to this research by developing semantic negotiation protocols and alignment mechanisms between vector spaces, while offering infrastructure and services adapted to this new era of distributed AI,” concludes the researcher.

This text has been translated by an artificial intelligence.