Over-the-Air Model Aggregation via MAC PDU Transmission
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current wireless communication systems face challenges in efficiently aggregating and updating AI models in federated edge learning systems due to limitations in radio resources and MAC/PHY layer compatibility, leading to degraded performance and slow model updates.
Innovation Solution
The implementation of a method where user equipment (UE) determines quantized parameters or gradients in a recurrent neural network (RNN) based on AI modeling and transmits these as a MAC PDU or set of bits on a physical uplink shared channel (PUSCH) that overlaps with other UEs' resources, allowing for central aggregation and updating of global models without channel encoding, facilitating over-the-air model aggregation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If quantized parameters or gradients are transmitted using traditional channel encoding methods, then reliability of transmission is improved, but productivity of model aggregation is degraded due to overhead and processing time
Solution Approach 1:
The patent extracts the essential information (quantized parameters or gradients) from the traditional communication protocol stack by directly using MAC PDUs or bit sets without higher-layer processing. This removes unnecessary protocol overhead and processing steps while maintaining the core transmission function, thereby improving productivity without sacrificing reliability.
Solution Approach 2:
The patent changes the transmission parameter format by using quantized parameters or gradients directly embedded in MAC PDUs or as bit sets, rather than traditional encoded data formats. This parameter change enables more efficient resource utilization and faster processing at the base station, improving model aggregation speed while maintaining transmission reliability through the structured format.
2Productivity
If overlapping PUSCH resources are used for multiple UEs, then productivity of model aggregation is improved by utilizing available resources, but device complexity increases due to resource coordination requirements
Solution Approach 1:
The patent merges the model aggregation transmissions from multiple UEs onto overlapping PUSCH resources. By combining these transmissions in the time-frequency domain and leveraging the base station's ability to separate and aggregate the quantized parameters, the system improves productivity by utilizing available radio resources more efficiently while the base station handles the coordination complexity centrally.
Solution Approach 2:
The patent makes the PUSCH resources universal by allowing them to serve multiple functions: traditional uplink data transmission and model aggregation transmission. The MAC PDU structure is designed to carry both regular uplink data and quantized model parameters, enabling the same physical resources to handle multiple types of traffic simultaneously, thereby improving resource utilization without requiring dedicated resources for model aggregation.
3Productivity
If channel encoding is disabled for quantized parameters, then productivity is improved by reducing processing overhead, but reliability may be degraded without error protection
Solution Approach 1:
The patent applies quantization as a preliminary action before transmission. By quantizing the model parameters or gradients beforehand at the UE side, the data is transformed into a format that is more robust to transmission errors and better suited for direct aggregation. This preliminary processing reduces the need for heavy channel encoding while maintaining reliability, as the quantized format inherently tolerates some transmission imperfections.
Data Source
AI summary
Various aspects of the present disclosure generally relate to wireless communication. In some aspects, a user equipment (UE) may determine quantized parameters in a recurrent neural network (RNN), or gradients to derive the RNN, based at least in part on artificial intelligence (AI) modeling at the UE as part of a federated edge learning system. The UE may generate a message that indicates the quantized parameters or gradients determined by the UE. The message may include a medium access control (MAC) protocol data unit or a set of bits obtained from a MAC layer, packet data convergence protocol layer, or application layer. The UE may transmit the message to a base station on a physical uplink shared channel (PUSCH) radio resource that overlaps with PUSCH radio resources used by other UEs. Numerous other aspects are provided.