METHODS FOR ENABLING OVER-THE-AIR FEDERATED LEARNING USING UPLINK PRE-EQUALIZATION

Smart overview of the Invention

The patent describes a method for implementing over-the-air federated learning (OTA-FL) on a wireless transmit/receive unit (WTRU). This involves receiving configuration information for training rounds of an AI/ML model, which includes allocations of over-the-air reference signal (OTA-RS) resources and mechanisms for uplink pre-equalization. The WTRU receives indications of uplink resource allocations to transmit pre-equalized, locally updated AI/ML model parameters during these training rounds.

Local Training and Pre-equalization

Local AI/ML model training is performed by the WTRU to determine a trained AI/ML model. During this process, local model parameters or gradients are generated for a federated learning task. Channel estimation is conducted using OTA-RSs, and uplink pre-equalization parameters are derived from this estimation. The locally updated model parameters are pre-equalized for transmission based on the determined channel estimation and uplink resource allocations.

Transmission and Resource Allocation

Pre-equalized locally updated AI/ML model parameters are transmitted over the allocated uplink resources during OTA-FL training rounds. The configuration information may include mapping instructions to associate these parameters with uplink resource elements. Incremental updates to the AI/ML model parameters are transmitted, representing differences from a reference model, which helps refine the global AI/ML model received by the WTRU.

Channel Estimation and Calibration

Channel estimation is enhanced using OTA-RSs with higher density than traditional Channel State Information Reference Signals (CSI-RSs) in both frequency and time domains. Calibration information from the network is received to adjust uplink phase and amplitude errors during OTA-FL training rounds. This ensures accurate transmission of pre-equalized parameters.

Dynamic Allocation and Coordination

Uplink resources are dynamically allocated to multiple WTRUs performing the same OTA-FL task. This allocation is managed via uplink scheduling grants with a group Radio Network Temporary Identifier (RNTI). The mechanism for uplink pre-equalization may involve phase correction, channel inversion, or truncated channel inversion. The network coordinates multiple OTA-FL training rounds to refine the global AI/ML model effectively.