Non-Coherent Orthogonal Modulation for OTA Federated Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In wireless communication systems, particularly in 5G NR, there is a challenge with analog over-the-air federated learning (FL) that requires channel pre-compensation at user equipment (UE) for coherent combining of signals, which is difficult without channel state information (CSI) and may exceed the UE's transmit power capabilities.
Innovation Solution
Implementing non-coherent orthogonal modulation for OTA FL gradient aggregation, where each UE estimates and quantizes its local gradient and transmits it using a non-coherent orthogonal modulation scheme on multiple resources, eliminating the need for channel pre-compensation and allowing simultaneous transmission without CSI.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If channel pre-compensation is implemented at UE for coherent combining, then signal combining coherence is improved, but UE transmit power requirements increase and device complexity increases
Solution Approach 1:
The patent replaces the mechanical/channel-based pre-compensation approach with a mathematical modulation approach. Instead of adjusting transmit signals based on channel state information (which requires power and complexity), the system uses orthogonal modulation codes (e.g., Hadamard codes) that are applied independently of channel conditions. This substitution eliminates the need for channel pre-compensation while maintaining coherent combining through the mathematical properties of orthogonal codes.
Solution Approach 2:
The patent changes the modulation parameter from channel-dependent amplitude/phase adjustment to channel-independent orthogonal code selection. By using orthogonal modulation codes as the key parameter instead of channel state information for signal shaping, the system achieves coherent combining without requiring power-intensive pre-compensation operations at the UE.
2Reliability
If channel pre-compensation is implemented at UE for coherent combining, then signal combining coherence is improved, but device complexity increases
Solution Approach 1:
The patent replaces the complex channel pre-compensation mechanism with a simpler orthogonal code modulation system. Instead of requiring the UE to estimate and compensate for channel effects (which involves complex signal processing), the system uses pre-defined orthogonal codes that are applied through simple code selection and multiplication operations, significantly reducing UE complexity.
Solution Approach 2:
The orthogonal modulation codes inherently provide the coherence needed for combining without requiring external channel pre-compensation assistance. The codes self-contained mathematical properties (orthogonality) automatically ensure coherent combining at the server side, eliminating the need for complex UE-side channel management operations.
3Reliability
If CSI is required for coherent combining, then signal combining performance is improved, but system adaptability decreases
Solution Approach 1:
The patent substitutes the CSI-dependent coherent combining mechanism with a CSI-independent orthogonal modulation approach. The system achieves signal combining performance through the mathematical orthogonality of codes rather than through channel state knowledge, thereby eliminating the adaptability constraint imposed by CSI requirements while maintaining combining performance.
4Productivity
If quantization is applied to gradient, then communication efficiency is improved, but gradient precision decreases
Solution Approach 1:
The patent segments the gradient communication into two distinct phases: local quantization at the UE (for compression and efficiency) and precise aggregation at the server (for maintaining accuracy). By separating these functions, the system achieves communication efficiency through quantization while preserving gradient precision through the server's aggregation operation that operates on the quantized values without further quantization loss.
Data Source
AI summary
Aspects of the disclosure are directed to quantized orthogonal modulation for non-coherent over-the-air (OTA) federated learning (FL). In some examples, a wireless device may compare a gradient to a threshold value to determine whether the gradient satisfies a threshold condition. The wireless device may also output, for transmission to a network node, signaling via one of a first resource or a second resource, wherein the signaling is outputted via the first resource if the gradient satisfies the threshold condition, and wherein the signaling is outputted via the second resource if the gradient does not satisfy the threshold condition.


