Power control in over the air aggregation for federated learning
A parameter server located at a base station may coordinate federated learning among multiple user equipment (UEs) using over-the-air (OTA) aggregation with power control to mitigate aggregation distortion due to amplitude misalignment. The parameter server may select a first group of UEs for a first OTA aggregation session of a federated learning round based on a common received power property of each UE in the first group of UEs. The parameter server may transmit a global model to the first group of UEs. Each UE in the first group may train the global model based on a local dataset and transmit values associated with the trained local model. The parameter server may receive, 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.
1. A method for federated learning at a UE, comprising:
receiving a global model from a parameter server;
training the global model based on training data to determine a plurality of local epochs;
performing truncated channel inversion based on a power headroom and a threshold for truncated channel inversion for each of a plurality of values associated with at least a last local epoch of the plurality of local epochs; and
transmitting an amplitude modulated analog signal to the parameter server for each respective value of the plurality of values, wherein a transmission power of the amplitude modulated analog signal is scaled based on the respective value associated with at least the last local epoch.
2. The method of claim 1 , wherein performing truncated channel inversion comprises truncating one or more of the plurality of values associated with at least the last local epoch to meet the power headroom in response to a transmit power after channel inversion exceeding a power headroom.
3. The method of claim 1 , further comprising receiving a maximum transmission power and the threshold for truncated channel inversion from the parameter server.
4. The method of claim 1 , further comprising determining the power headroom and the threshold for truncated channel inversion based on a tradeoff between transmit power penalty and value accuracy.
5. The method of claim 1 , further comprising determining the power headroom and the threshold for truncated channel inversion as a function of pathloss and a number of training samples configured by the parameter server.
6. The method of claim 1 , further comprising choosing a configured uplink resource for transmitting the amplitude modulated analog signal based on a downlink pathloss.
7. The method of claim 1 , further comprising scaling the transmission power based on a confidence level of a gradient estimate after a last local epoch.
8. An apparatus for federated learning at a user equipment (UE), comprising:
a memory storing computer-executable instructions; and
at least one processor communicatively coupled with the memory and configured to execute the instructions to:
receive a global model from a parameter server;
train the global model based on training data to determine a plurality of local epochs;
perform truncated channel inversion based on a power headroom and a threshold for truncated channel inversion for each of a plurality of values associated with a least a last local epoch of the plurality of local epochs; and
transmit an amplitude modulated analog signal to the parameter server for each respective value of the plurality of values, wherein a transmission power of the amplitude modulated analog signal is scaled based on the respective value associated with at least the last local epoch.
9. The apparatus of claim 8 , wherein the at least one processor is configured to truncate one or more of the plurality of values associated with at least the last local epoch to meet the power headroom in response to a transmit power after channel inversion exceeding a power headroom.
10. The apparatus of claim 8 , wherein the at least one processor is configured to receive a maximum transmission power and the threshold for truncated channel inversion from the parameter server.
11. The apparatus of claim 8 , wherein the at least one processor is configured to determine the power headroom and the threshold for truncated channel inversion based on a tradeoff between transmit power penalty and value accuracy.
12. The apparatus of claim 8 , wherein the at least one processor is configured to choose a configured uplink resource for transmitting the amplitude modulated analog signal based on a downlink pathloss.
13. The apparatus of claim 8 , wherein the at least one processor is configured to determine the power headroom and the threshold for truncated channel inversion based on a function of pathloss and a number of training samples configured by the parameter server.
14. The apparatus of claim 8 , wherein the at least one processor is configured to scale the transmission power based on a confidence level of a gradient estimate after a last local epoch.
15. An apparatus for federated learning at a user equipment (UE), comprising:
means for receiving a global model from a parameter server;
means for training the global model based on training data to determine a plurality of local epochs;
means for performing truncated channel inversion based on a power headroom and a threshold for truncated channel inversion for each of a plurality of values associated with at least a last local epoch of the plurality of local epochs; and
means for transmitting an amplitude modulated analog signal to the parameter server for each respective value of the plurality of values, wherein a transmission power of the amplitude modulated analog signal is scaled based on the respective value associated with at least the last local epoch.
16. The apparatus of claim 15 , wherein the means for performing truncated channel inversion is configured to truncate one or more of the plurality of values associated with at least the last local epoch to meet the power headroom in response to a transmit power after channel inversion exceeding a power headroom.
17. The apparatus of claim 15 , further comprising means for receiving a maximum transmission power and the threshold for truncated channel inversion from the parameter server.
18. The apparatus of claim 15 , further comprising means for determining the power headroom and the threshold for truncated channel inversion based on a tradeoff between transmit power penalty and value accuracy.
19. The apparatus of claim 15 , further comprising means for determining the power headroom and the threshold for truncated channel inversion as a function of pathloss and a number of training samples configured by the parameter server.
20. The apparatus of claim 15 , further comprising means for choosing a configured uplink resource for transmitting the amplitude modulated analog signal based on a downlink pathloss.
21. The apparatus of claim 15 , further comprising means for scaling the transmission power based on a confidence level of a gradient estimate after a last local epoch.
22. A non-transitory computer-readable medium storing computer executable code, the code when executed by a processor causes the processor to:
receive a global model from a parameter server;
train the global model based on training data to determine a plurality of local epochs;
perform truncated channel inversion based on a power headroom and a threshold for truncated channel inversion for each of a plurality of values associated with at least a last local epoch of the plurality of local epochs; and
transmit an amplitude modulated analog signal to the parameter server for each respective value of the plurality of values, Wherein a transmission power of the amplitude modulated analog signal is scaled based on the respective value associated with at least the last local epoch.
23. The non transitory computer-readable medium of claim 22 , wherein the code when executed by a processor causes the processor to truncate one or more of the plurality of values associated with at least the last local epoch to meet the power headroom in response to a transmit power after channel inversion exceeding a power headroom.
24. The non-transitory computer-readable medium of claim 22 , wherein the code when executed by a processor causes the processor to receive a maximum transmission power and the threshold for truncated channel inversion from the parameter server.
25. The non-transitory computer-readable medium of claim 22 , wherein the code when executed by a processor causes the processor to determine the power headroom and the threshold for truncated channel inversion based on a tradeoff between transmit power penalty and value accuracy.
26. The non-transitory computer-readable medium of claim 22 , wherein the code when executed by a processor causes the processor to choose a configured uplink resource for transmitting the amplitude modulated analog signal based on a downlink pathloss.
27. The non-transitory computer-readable medium of claim 22 , wherein the code when executed by a processor causes the processor to determine the power headroom and the threshold for truncated channel inversion based on a function of pathless and a number of training samples configured by the parameter server.
28. The non-transitory computer-readable medium of claim 22 , wherein the code when executed by a processor causes the processor to scale the transmission power based on a confidence level of a gradient estimate after a last local epoch.