As telecommunications networks evolve toward 6G, a key technical question is where artificial intelligence should run, how network software can remain interoperable, and how AI systems can learn across distributed environments without requiring all raw data to be centralized. Ravi Shanker’s research examines these challenges through a proposed architecture combining federated learning, split learning, and Low-Rank Adaptation (LoRA).
Shanker’s work draws on experience supporting designated markets during AT&T’s Nokia-to-Ericsson radio access network transition through Verveba Telecom, where he served as a Command Control Center lead. His responsibilities included validating technical and configuration information before maintenance windows, coordinating implementation activities, tracking exceptions, and monitoring network performance following integration. That operational experience provides context for his research into distributed network intelligence.
Reliable automation depends on accurate configuration data, clearly defined responsibilities, consistent workflow states, and the ability to identify and resolve exceptions. These principles are also relevant to networks designed to support increasingly automated operations.
In his paper, “AI-Driven Federated and Split Learning Frameworks for Optimized 6G Network Architectures,” published in the 2025 International Conference on Computing Technologies & Data Communication proceedings and indexed by IEEE Xplore, Shanker proposes a theoretical framework for distributing AI computation across devices, edge infrastructure, and centralized data centers. The framework considers how computing tasks could be allocated according to latency, memory, energy consumption, network conditions, and user mobility. Rather than requiring every AI task to travel to a central cloud, the proposed architecture considers which functions could be performed on devices or nearby edge nodes and which would require more powerful centralized resources.
Federated learning is one component of the framework. It allows participating devices or systems to train models locally and exchange model updates rather than directly pooling their raw data. In distributed wireless networks, this approach could support collaborative learning across devices and edge infrastructure while reducing the need to transfer sensitive information. However, communication overhead, inconsistent device capabilities, synchronization, and potential information leakage remain important technical challenges.
Split learning offers another approach by dividing a model’s computation across multiple systems. A device can process an initial portion of a model while an edge server or cloud system handles subsequent layers. This can reduce the computational burden on constrained devices, although the division introduces dependencies between systems and requires careful management of latency, intermediate data, and mobility.
The third component, Low-Rank Adaptation, enables selected parameters of a pretrained model to be adapted without updating the entire model. Shanker’s framework considers how compact model adapters could support task-specific or domain-specific behavior while larger shared models remain on edge or cloud infrastructure.
The proposed architecture combines these methods to explore how distributed collaboration, model partitioning, and parameter-efficient adaptation might work together. Its practical implementation would require orchestration mechanisms capable of determining where models and adapters should reside, how they should move as users change network locations, and how performance, privacy, and reliability should be maintained.
The research also provides a technical perspective on broader discussions about open wireless architecture and converged connectivity. Open interfaces could support interoperability among network equipment, edge computing resources, and AI services, while fiber, wireless, and satellite infrastructure could contribute different forms of transport and coverage. Coordinating these resources with computing capacity and model placement presents a further challenge for future network design.
Shanker’s paper presents a theoretical framework rather than a production-validated 6G implementation. It does not establish that the architecture has been deployed by AT&T, nor does it represent an official AT&T network design. Instead, it explores technical questions that may inform future work on distributing intelligence across wireless and computing infrastructure.
The relationship between network modernization and distributed AI is central to the framework: operational systems require observable states, accountable exception handling, and reliable feedback, while AI-native networks must also determine where learning and inference occur and how automated decisions are governed. Further engineering and validation would be needed to establish the framework’s performance, security, and suitability for real-world deployment.
