Communication Engineering in the Era of Generative AI
Two scenes from the future
In one scene, a group of football fans watches the championship finals between teams A and B on a screen. It is a live broadcast. Their team A scores and a chaotic celebration is ignited. Another group of football fans, supporting team B, watches the same game through a live broadcast, but in that broadcast the team B scores. Two opposite outcomes are celebrated simultaneously. I leave the consequences of this to the reader’s imagination and would only mention here that an early indication of the upcoming war in Yugoslavia was hinted in a football match. Having two opposite outcomes streamed live is possible because one of the live broadcasts, or even both of them, uses a communication link that has been hijacked by a generative Artificial Intelligence (AI). This would be a generative AI specialized for creating video content in real time, an evolutionary offspring of ChatGPT or other presently available generative AIs.
In a second scene, a humanoid robot and an autonomous vehicle are equipped with multiple communication interfaces. There is nothing futuristic in that fact, as an average smartphone in 2023 has a significant number of interfaces, such as several types of cellular connections (2G to 5G), WiFi, Bluetooth, or NFC. The futuristic element in this scene is that the robot and the vehicle use a generative AI to evolve a communication protocol that is capable to use the existing communication interfaces and hardware in ways that are understood only by two of them. The protocol generated in this way is resilient, as it skillfully uses all interfaces upon jamming and is well-integrated with the coordination/movement of the robot and the vehicle. In short, the robot and the vehicle make a distributed entity that is capable to manoeuvre resiliently among people and machines.
The two scenes from above indicate two principle ways in which generative AI affects communication systems. The first scene shows how generative AI can affect the data transported through the communication systems. The second scene reflects possible ways in which the generative AI can have an impact to the communication system design and communication protocol.
Predictive Generation replaces Communication
In the extreme case, the use of AI completely removes the need for communication. For example, Alice wants to tell something to Bob, either through voice, text, or multimedia. If Bob has an AI module that can make a perfect prediction about what Alice wants to tell, the module can just generate the content locally for Bob. Indeed, the very basic definition of information as a measurement of uncertainty or surprise implies that what can be predicted does not need to be communicated.
It gets more interesting if Bob cannot perfectly predict what Alice wants to tell. In that case, Bob’s generative AI module can create content that approximates what Alice would have told him. Generating approximate responses in order to save communication bandwidth is a technique that has already been considered in approximate data collection from wireless sensor networks. The main idea is that the access point has a certain model for the data provided by the sensor network. Upon arrival of a query, say from the cloud, for data collection from the sensors, the access point can opt not to communicate with the sensors if it has a very confident prediction of the sensory data. Thus, the access point sends back to the cloud a prediction, an estimate, rather than the actual value collected from the sensors. Sending approximate, but meaningful information in order to save communication bandwidth is often put in the context of semantic communications.
In the extreme case, replacing communication with prediction and generation leads to a dystopian scenario: no one actually communicates with anyone else because everything can be predicted or, at most, acceptably approximated. Besides directly cancelling the possibility for existence of a free will, this extreme scenario is not possible due to at least two reasons. First, the physical world is essentially unpredictable and the uncertainty about the state of some device or entity only increases with time. These observations are formalized through the quantum mechanic principles and the second law of thermodynamics. Second, not every device will have the computational power required to store large AI models, such that the capability for prediction and generation will not be uniform across the devices. In fact, it is highly likely that the distribution of the AI capabilities will be highly unequal, with relatively few clusters/clouds with extremely high computational power and capability for generative AI.
Let us look into communication over links in which the interconnected nodes have unequal generative AI capabilities. To put things in perspective, consider a link between node X, with high generative AI capability, and node Y with low AI capabilities. Let us call this a setup with extreme AI inequality. Thus, node X can be thought of having a significant knowledge about what Y wants to tell, such that Y needs to send only “a little” data over the communication link to X. The knowledge that X has about Y is called a side information in information/communication theory. There are already ingenious communication methods for distributed source coding, such as Slepian-Wolf or Wyner-Ziv coding, and similar ideas can be used to design schemes that allow Y to tell only a little to X without Y even knowing exactly what X knows. Looking at the communication from X to Y, since Y has a limited or even no capability to make predictive generation of the content sent by X, such that X needs to communicate to Y as usual. In other words, the generative AI has little or no effect on the link from X to Y. As an example of this setup with extreme AI inequality, X can be cloud/edge node and Y can be a simple sensor or actuator.
The cases that are much more interesting than the case with extreme AI inequality are the ones with moderate AI inequality, where both X and Z (see figure) have nontrivial AI capabilities. Here the communication can start by having X and Z exchange Machine Learning/Neural network models that will initialize generative AI capabilities at both nodes that suitable for the type of communication that will take place. The objective is to initialize the generative AI capability in a way that will decrease the need for communication; updates are sent only when the generative model deviates from the actual information to be sent. There are already works on neural network transmission over wireless links and, in the context of conventional communication, one can understood this phase of the communication protocol as initialization during which the nodes agree upon the compression method and parameters.
In his insightful article “ChatGPT Is a Blurry JPEG of the Web”, the sci-fi author Ted Chiang starts with an example in which statistical compression of data leads to undesired results, since it does not directly account for the importance/semantics of certain data in a certain context. This problem becomes very relevant and acute when the communication nodes rely on generative AI modules. Namely, the considerations above, both for the setup with extreme AI inequality and the one with moderate AI inequality, point towards the use of compression methods for the purpose of reducing the required communication bandwidth. The compression should adapt to two aspects:
(1) The importance that is given to a given type of information by the recipient. For instance, the quantization should not blur parts of an image that are of vital importance for the receiver.
(2) The fact that the communicating nodes are continuously learning, which would require continuous updates of the generative AI models.
Hallucinating about Communication Protocols
So far we have mostly discussed how generative AI can affect the data sent over the communication links. However, exchanging generative models prior to the communication is a step that shows how generative AI can change the design of the communication protocols, or even entire communication systems. Going a step further, one can use generative AI, tailored for designing communication protocols, to generate proposals for protocols. These protocols may in turn rely upon generative AI modules, tailored for compression of the data that needs to be communicated. Thus, the protocol-generating AI designs a protocol that works with data-generating AIs in order to make the communication more efficient.
As another example, generative AI can be used to design protocols that operate and evolve under the assumption that the knowledge/intelligence at the communicating nodes continuously increases. One can think of it as some form of evolutionary communication protocols, whose elements in terms or signaling or data transmission formats evolve continuously with the network topology, services, and intelligence at the interconnected nodes.
Generative AI can be used to design communication protocols that are too complex to be created by a human engineer. For instance, rather than specifying the control signaling information in a way in which “we have always done it”, generative AI may come up with alternative signaling schemes that appear to be better suited for the scenario at hand. Here signaling information can include e.g. synchronization sequences, metatada for packets, and similar. These protocols are result of AI hallucinations as they are not based on already existing protocols, which would serve as input data, but protocols generated in the “mind” of generative AI. As described in scene two from the beginning of this article, protocols generated through generative AI hallucinations may go rogue and create a communication system that operates outside the intended boundaries, potentially leading to unintended consequences for the humans and other nodes on the network.
Yet, having a playground for controlled hallucinations can lead to creative and efficient design of communication protocols. One appealing idea for setting the design playground would be to let the large language models, such as ChatGPT, go through the protocol specifications, such as 3GPP or IEEE 802 11, and propose efficient implementation or point out inefficiencies and the protocols, along with solutions how to address them [1] [2]. Using generative AI for designing communication protocols will shift the required engineering skills towards skills for creating the playground for controlled protocol hallucinations, similar to prompt engineering for chatbots. This includes setting the operating conditions for the protocol, ask counterfactual questions by stress-testing the protocols with extreme scenarios, and similar. In addition, the communication engineers will need to extract explanations about the actions taken by the protocol; that is, there can be an explainable AI interface between the human protocol/network designers and the AI modules.
Access links as Guardians of Truth
In a recent article “Surprise-Inspired Networking”, David Tennenhouse offers a refreshing view on the evolution on networking: the new data and the semantic surprises will come from the network edge rather from the network core or the cloud. This necessitates new communication architectures and designs, tailored to the data that comes from the edge, while relying on the computational and AI power of the cloud serves and other capable nodes at the network core.
Bringing generative AI to this setup brings a different perspective on the data coming from the access links and the edge. The devices connected to the edge and their access links are responsible to inject truthful data to the network, which represents the information and the events in the real, physical world. The wireless access links, as well as the sensory and positioning data, will be put in a position of guardians of truth, in contrast to the synthetic data that can be produced and injected in the network by a generative AI (recall the football match example from the introduction). This would require truthfulness of the communication and sensory devices as well as protocols that will guarantee the data/information provenance; otherwise, the network and its users will be flooded by synthetic data and severely distorted truth, leading to unintended consequences. Truthfulness of the wireless access offers an additional raison d’être for 6G wireless communication systems, as there is a consensus that 6G will integrate three modalities for information exchange and acquisition: digital communication, sensors, and positioning/localization.
We can speculate about an emergence of a completely new set of wireless services, intended to gather data from the truthful access links and sensors in order to fight, in real-time, the disinformation created by powerful generative AI. The whole network will be running a plethora of Turing-like tests to detect real vs. synthetic data. At a large scale, the network may establish real-time data markets, in which economic incentives are used to promote transmission, storage, and processing of truthful data, rather than synthetic data. Genuine data should have a higher market value as compared to the data created by generative AI, while the providers of the truthful data should be suitably incentivized. The widespread use of AI has already brought some interesting ideas for data valuation [3]; the context of real-time generative AI vs. truthful data calls opens up for new set of challenges. Although blockchains lost some of their popularity in the general public due to the low performance of the cryptocurrencies, technologies based on distributed ledgers and distributed consensus are likely to grow in importance as the real-time battle between true and fake data intensifies.
Summary and Outlook
The growing intelligence in the interconnected nodes an, specifically, generative AI, can have a profound impact on the way communication systems are designed and operated, as it affects both the data exchanged through the network as well as the design of the protocols used to operate the network. Rather than statistical compression of bits and their faithful transmission, generative AI will drive the future communication network towards communication that resembles the theory of mind, in which the network nodes and devices try take the perspectives of their communication peers, infer and predict their intentions and actions, or establish a shared knowledge for a more efficient communication. Communication engineers will need skills to generate, test, explain, and control protocols that are hallucinated by a generative AI. Following the launch of Apple’s Vision pro as the first spatial computer, the fusion of physical and digital world becomes more likely to occur at a large scale within the coming period. In this context, the future network will include AI-capable nodes that can generate synthetic data in real-time; this puts the wireless mobile devices, along with their access links and sensors, in a position of guardians of the truth about the physical world.
References
[1] Prof. Osvaldo Simeone and Dr. Onur Shahin, private communication.
[2] Merouane Debbah, keynote talk at IEEE ICC 2023, https://icc2022.ieee-icc.org/program/keynotes.html#Merouane and https://arxiv.org/pdf/2306.07933.pdf.
[3] Anish Agarwal, Munther Dahleh, and Tuhin Sarkar. 2019. A Marketplace for Data: An Algorithmic Solution. In Proceedings of the 2019 ACM Conference on Economics and Computation (EC ‘19). Association for Computing Machinery, New York, NY, USA, 701–726. https://doi.org/10.1145/3328526.3329589
