US 12,445,979 B2
Power control in over the air aggregation for federated learning
Eren Balevi, Brooklyn, NY (US); Taesang Yoo, San Diego, CA (US); Tao Luo, San Diego, CA (US); Srinivas Yerramalli, San Diego, CA (US); Junyi Li, Greentown, PA (US); and Hamed Pezeshki, San Diego, CA (US)
Assigned to QUALCOMM Incorporated, San Diego, CA (US)
Filed by QUALCOMM Incorporated, San Diego, CA (US)
Filed on Jun. 3, 2024, as Appl. No. 18/732,407.
Application 18/732,407 is a division of application No. 17/544,581, filed on Dec. 7, 2021, granted, now 12,035,256.
Prior Publication US 2024/0323870 A1, Sep. 26, 2024
Int. Cl. H04W 52/50 (2009.01); H04W 4/06 (2009.01); H04W 52/22 (2009.01); H04W 52/24 (2009.01); H04W 52/36 (2009.01)
CPC H04W 52/50 (2013.01) [H04W 4/06 (2013.01); H04W 52/225 (2013.01); H04W 52/242 (2013.01); H04W 52/36 (2013.01)] 19 Claims
OG exemplary drawing
 
1. A method for federated learning at a base station, comprising:
selecting a first group of user equipment (UEs) for a first over-the-air (OTA) aggregation session of a federated learning round based on a common received power property of each UE in the first group of UEs;
transmitting a global model to the first group of UEs;
configuring at least one UE of the first group of UEs with a rule for determining a maximum transmission power and a threshold for truncated channel inversion as a function of a pathloss and a number of training samples; and
receiving, on resource elements for the first group of UEs, a first aggregate amplitude modulated analog signal representing a combined response from the first group of UEs, the combined response being a sum of respective values associated with trained local models for an aggregation period determined at each UE based on the global model and a local dataset.