.
Generative AI (Gen-AI) has witnessed a remarkable year, marked by a surge in interest and widespread adoption across new domains and user bases. We have seen the rapid emergence of notable large language models (LLMs) including OpenAI’s ChatGPT/GPT-4 and Meta’s Llama-2 as well as powerful generative text to image diffusion models like OpenAI’s DALL-E 2, HuggingFace’s Stable Diffusion XL, and Midjourney. These models demonstrated their impressive ability to take vast amounts of information from billions of documents and examples and turn it into concise representations. They can then provide detailed natural language responses and create realistic images reflecting complex scene prompts.
The disruption of Gen-AI has been impactful in current use cases thus far, but we have only seen the beginning of the impact of the technology. Natural language and image processing have seen significant changes, as they can now efficiently search, summarize, provide detailed answers, even style text, create poetry, lyrics, software, various forms of language content, and images. This is only the beginning of a long list of information, data processing applications and industries which can broadly benefit through the adoption of Generative AI. Many fields are set to undergo significant transformations and acceleration, among these industries being wireless communications – a critical field at the foundation for modern connectivity, industry, collaboration, entertainment, and daily functions.
Generative AI is fundamentally about distilling large volumes of data examples into (relatively) compact and accurate models, to then be used to generate representative new examples. Unlike many other AI methods that rely on supervision in the form of labeled data, Generative AI can operate on raw unlabeled data. This is accomplished by learning to predict data from other data, for instance next tokens or words, pixels, voxels, or other data points. But it may also be used for a wide range of physical distributions which are hard to model such as material properties, particle behaviors, light and EM wave propagation models, and similar based on measurement data. In doing so, Generative AI can in some cases better model complex data and physical distributions and phenomena from unstructured data incredibly well, and more accurately than any known closed form representations. Wireless communications in particular is one field where more accurate models for the world are needed, and can greatly impact performance.
Leveraging Generative AI in Wireless Systems
Machine learning, especially deep learning, has two primary roles in improving wireless communication systems. First, it helps capture real-world effects and behavior in wireless systems more accurately, enabling better optimization for real-world conditions. Second, machine learning models are directly applied to solve issues like signal transmission, reception, and error correction. These models are fine-tuned to achieve specific objectives such as increasing capacity, reducing latency, improving throughput, and minimizing power consumption. DeepSig was founded to address both challenges effectively, offering mature software solutions, components, and IP for various wireless systems. Generative AI plays a crucial role in the first aspect by modeling realistic wireless environments and effects, which in turn is used by the second. This year, interest in Gen AI for wireless has grown significantly – for example, excitement around the use of LLMs in wireless infrastructure in 5G and 6G systems was palpable at the recent Machine Learning for Wireless Communications ‘AI For Good’ workshop [1] held by the United Nations (UN), International Telecommunications Union (ITU). Here several use cases for LLMs were highlighted ranging from customer support chat-bots and knowledge management systems to code synthesis and letting LLMs perform network parameter tuning.
DeepSig’s focus on Gen-AI, as discussed at the ITU event, extends beyond LLMs. It includes optimizing 5G vRAN infrastructure, next-generation 5G Advanced, and 6G networks. This focus also encompasses data collection and training generative models for various non-linguistic physical-world wireless phenomena. Among these includes accurately modeling things like wireless signals, hardware effects, user behavior, interference, mobility and antenna patterns, and various other complex factors that directly impact wireless performance.
Building Gen-AI Beyond Linguistic Systems for Wireless at DeepSig
Generative deep learning models for physical world wireless phenomena have held significant promise for the past few years since their capacity for modeling complex data distributions at scale was first demonstrated. One of the first areas we looked at this technology for at DeepSig was for modeling the channel response (i.e. impulse response) of a wireless channel, with hardware effects and impairments included. Doing this in a closed form mathematical expression has always been a somewhat difficult problem – but there have been many well-known wireless channel models proposed and used over the years. Among these, current standardized time delay line (TDL) and cluster delay line (CDL) closed-form models are widely used today within 5G wireless system test and performance measurement and have come a long way in more closely representing the real-world. However, they are still vastly more simplified as compared with the real world in which communications systems are deployed. Closing this gap of channel model deficit holds significant opportunity in terms of better capacity, resilience, spectrum utilization, spectrum sharing, efficiency, etc.
DeepSig has been investigating Gen-AI models for these purposes for a number of years, having published several papers on the topic nearly 5 years ago [3][4] and having several patents issued on the topic including US 10531415 and US 11259260. Throughout these efforts, DeepSig has built several deep generative models for channel approximation leveraging techniques including generative adversarial networks (GANs) as well as variational autoencoders (VAEs) and other methods effectively learn how to model complex real world channel impulse responses. These can be a powerful tool when training machine learning based wireless systems such as DeepSig’s OmniPHY5G wireless uplink (PUSCH) receiver solution, to perform in the real world accurately and effectively. Similarly, ML driven RF sensing as we have developed with OmniSIG, allows for broad awareness of out-of-sector emitters such as adjacent sectors, adjacent networks, unauthorized emissions, jamming or unintended emissions, or otherwise – providing volumes of data on both propagation data as well as emitter activity offers the raw information which generative models and communications systems can use to provide significantly improved performance.
Volumes of data on localized propagation and channel distributions, as well as emitter activity over time and space, may then be captured through both data collection campaigns, and ultimately through infrastructure and mobile devices themselves to provide a complete and continuous picture of the physical wireless world.
Recently, the idea of a “Digital Twin” has been popularized as a method to capture and represent physical world activity within a digital equivalent which can be used for simulation, optimization, planning, or other processing or analysis. Both data collection and Gen-AI are critical enabling technologies in enabling the Wireless Digital Twin which can drive the next generation of air-interface performance, spectral efficiency, wireless system tuning, and spectrum sharing and orchestration. Numerous aspects of wireless systems can be tuned for local conditions using these generative models, including schedulers, resource allocation, modulation schemes, multi-antenna (MIMO) schemes, multi-access (NOMA) schemes, coding and piloting schemes, and a wide range of other aspects critical to wireless performance.
Accelerating Performance in OpenRAN with Generative AI
NTIA Public Wireless Supply Chain Innovation (PWSCI) Fund [2] recently selected DeepSig as the first commercial recipient of the grant to leverage and extend Generative AI tools within the 5G OpenRAN test and evaluation (T&E) ecosystem subsequently improving real-world performance. Under this grant, DeepSig is rapidly expanding data capture and Generative AI tools and datasets for critical testing and evaluation of the air interface under real-world wireless conditions which go beyond the standard 3GPP test scenarios. The objective is to enhance wireless system efficiency, coverage, security, and resilience. Moreover, these AI models and datasets contribute to training and optimizing various components, including OmniPHY®-5G, DeepSig’s AI-Native 5G receiver software, and will be leveraged within our wireless digital twin components and capabilities which are essential for competitive 5G Advanced and 6G wireless systems.
Generative AI has gained significant attention recently perhaps due to the popularization of ChatGPT, and at DeepSig, we’re thrilled about its impact on language models. Specifically, the broad range of use cases enabled and the potential to revolutionize the physical-world wireless domain through improved modeling, optimization, T&E, and digital twin applications for ORAN 5G, 5G Advanced, 6G and other applications.
We are accelerating our efforts into this space under the new NTIA led grant effort and with Open RAN ecosystem partners. We are excited to share our results and tools capabilities with the ecosystem to help advance wireless research, development, products, and intellectual property in the United States and its infrastructure partners.
Please check out our open positions if you’re interested in working with our team as we rapidly bring this technology to broad commercial deployments, DeepSig is hiring talented engineers, developers, and technologists who share this vision. See our posted openings here. Or, reach out to us at info@deepsig.ai today!
References:
- Machine Learning for Wireless Communications / AI For Good workshop
- NTIA Public Wireless Supply Chain Innovation (PWSCI) Fund grant
- Physical layer communications system design over-the-air using adversarial networks
- Approximating the void: Learning stochastic channel models from observation with variational generative adversarial networks
