Gradient dropping for federated learning
A UE may identify a plurality of local model update elements associated with an updated local machine learning model. The updated local machine learning model may have been generated based on a global machine learning model received from a base station and a local dataset. The UE may identify one or more local model update elements of the plurality of local model update elements for update element dropping based on at least one of a channel gain, a PAPR specification, an RF emission specification, or an RF condition. The UE may transmit, to the base station over a multiple access channel via analog signaling, at least some local model update elements of the plurality of local model update elements based on dropping of the identified one or more local model update elements.
1. An apparatus for wireless communication at a user equipment (UE), comprising:
a memory; and
at least one processor coupled to the memory and configured to:
identify a plurality of local model update elements associated with an updated local machine learning model, the updated local machine learning model having been generated based on a global machine learning model received from a base station and a local dataset, the plurality of local model update elements including a plurality of gradient elements or a plurality of updated local model parameters;
identify one or more local model update elements of the plurality of local model update elements for update element dropping based on at least one of a channel gain, a peak-to-average-power ratio (PAPR) specification, a radio frequency (RF) emission specification, or an RF condition; and
transmit, to the base station over a multiple access channel via analog signaling, at least some local model update elements of the plurality of local model update elements based on dropping of the identified one or more local model update elements, the analog signaling being associated with over-the-air (OTA) aggregation.
2. The apparatus of claim 1 , wherein the dropping of the identified one or more local model update elements is associated with one or more resource units, and each resource unit in the one or more resource units is one of a resource element (RE), a subcarrier, a resource block (RB), or a subband.
3. The apparatus of claim 2 , the at least one processor being further configured to:
transmit, to the base station, an indication of the one or more resource units associated with the dropping of the identified one or more local model update elements.
4. The apparatus of claim 3 , wherein the indication and the one or more resource units are associated with a same resource unit granularity.
5. The apparatus of claim 3 , wherein the indication is associated with a coarser resource unit granularity than the one or more resource units.
6. The apparatus of claim 1 , wherein the dropping of the identified one or more local model update elements is associated with all resource elements (REs) in a resource block (RB).
7. The apparatus of claim 1 , wherein the one or more local model update elements for update element dropping are identified based on a first channel gain associated with one or more resource units associated with the identified one or more local model update elements being less than a channel gain threshold.
8. The apparatus of claim 7 , the at least one processor being further configured to:
identify the channel gain threshold based on a UE transmission power budget.
9. The apparatus of claim 7 , the at least one processor being further configured to:
receive, from the base station, an indication of the channel gain threshold, the indication of the channel gain threshold corresponding to a scalar number or a formula.
10. The apparatus of claim 9 , wherein the indication of the channel gain threshold corresponds to the formula, and the at least one processor is further configured to identify the channel gain threshold based on the formula.
11. The apparatus of claim 7 , wherein the one or more resource units are associated with more than one spatial layer associated with a multiple-input-multiple-output (MIMO) transmission.
12. The apparatus of claim 7 , wherein a transmission power constraint associated with the UE is met based on the dropping of the identified one or more local model update elements.
13. The apparatus of claim 1 , wherein the one or more local model update elements for update element dropping are identified based on at least one of the PAPR specification, the RF emission specification, or the RF condition, and the identified one or more local model update elements are associated with one or more edge resource blocks (RBs) or one or more edge resource elements (REs).
14. The apparatus of claim 1 , further comprising a transceiver coupled to the at least one processor, the at least one processor being further configured to:
receive, from the base station, an indication of an upper fraction limit associated with the one or more local model update elements for update element dropping in the plurality of local model update elements.
15. A method of wireless communication at a user equipment (UE), comprising:
identifying a plurality of local model update elements associated with an updated local machine learning model, the updated local machine learning model having been generated based on a global machine learning model received from a base station and a local dataset, the plurality of local model update elements including a plurality of gradient elements or a plurality of updated local model parameters;
identifying one or more local model update elements of the plurality of local model update elements for update element dropping based on at least one of a channel gain, a peak-to-average-power ratio (PAPR) specification, a radio frequency (RF) emission specification, or an RF condition; and
transmit, to the base station over a multiple access channel via analog signaling, at least some local model update elements of the plurality of local model update elements based on dropping of the identified one or more local model update elements, the analog signaling being associated with over-the-air (OTA) aggregation.
16. An apparatus for wireless communication at a base station, comprising:
a memory; and
at least one processor coupled to the memory and configured to:
receive, from a user equipment (UE) over a multiple access channel via analog signaling, at least some local model update elements of a plurality of local model update elements based on dropping of one or more local model update elements of the plurality of local model update elements, the plurality of local model update elements being associated with an updated local machine learning model at the UE, the updated local machine learning model being based on a global machine learning model from the base station and a local dataset at the UE, the plurality of local model update elements including a plurality of gradient elements or a plurality of updated local model parameters, the one or more local model update elements for update element dropping being based on at least one of a channel gain, a peak-to-average-power ratio (PAPR) specification, a radio frequency (RF) emission specification, or an RF condition, the analog signaling being associated with over-the-air (OTA) aggregation; and
update the global machine learning model based on the received at least some local model update elements of the plurality of local model update elements.
17. The apparatus of claim 16 , wherein the dropping of the one or more local model update elements is associated with one or more resource units, and each resource unit in the one or more resource units is one of a resource element (RE), a subcarrier, a resource block (RB), or a subband.
18. The apparatus of claim 17 , the at least one processor being further configured to:
receive, from the UE, an indication of the one or more resource units associated with the dropping of the one or more local model update elements.
19. The apparatus of claim 18 , wherein the indication and the one or more resource units are associated with a same resource unit granularity.
20. The apparatus of claim 18 , wherein the indication is associated with a coarser resource unit granularity than the one or more resource units.
21. The apparatus of claim 16 , wherein the dropping of the one or more local model update elements is associated with all resource elements (REs) in a resource block (RB).
22. The apparatus of claim 16 , wherein the one or more local model update elements for update element dropping are based on a first channel gain associated with one or more resource units associated with the one or more local model update elements being less than a channel gain threshold.
23. The apparatus of claim 22 , wherein the channel gain threshold is based on a UE transmission power budget.
24. The apparatus of claim 22 , the at least one processor being further configured to:
transmit, to the UE, an indication of the channel gain threshold, the indication of the channel gain threshold corresponding to a scalar number or a formula.
25. The apparatus of claim 22 , wherein the one or more resource units are associated with more than one spatial layer associated with a multiple-input-multiple-output (MIMO) transmission.
26. The apparatus of claim 16 , wherein the one or more local model update elements for update element dropping are based on at least one of the PAPR specification, the RF emission specification, or the RF condition, and the one or more local model update elements are associated with one or more edge resource blocks (RBs) or one or more edge resource elements (REs).
27. The apparatus of claim 16 , the at least one processor being further configured to:
transmit, to the UE, an indication of an upper fraction limit associated with the one or more local model update elements for update element dropping in the plurality of local model update elements.
28. The apparatus of claim 16 , further comprising a transceiver coupled to the at least one processor.
29. A method of wireless communication at a base station, comprising:
receiving, from a user equipment (UE) over a multiple access channel via analog signaling, at least some local model update elements of a plurality of local model update elements based on dropping of one or more local model update elements of the plurality of local model update elements, the plurality of local model update elements being associated with an updated local machine learning model at the UE, the updated local machine learning model being based on a global machine learning model from the base station and a local dataset at the UE, the plurality of local model update elements including a plurality of gradient elements or a plurality of updated local model parameters, the one or more local model update elements for update element dropping being based on at least one of a channel gain, a peak-to-average-power ratio (PAPR) specification, a radio frequency (RF) emission specification, or an RF condition, the analog signaling being associated with over-the-air (OTA) aggregation; and
updating the global machine learning model based on the received at least some local model update elements of the plurality of local model update elements.
30. The method of claim 29 , wherein the dropping of the one or more local model update elements is associated with one or more resource units, and each resource unit in the one or more resource units is one of a resource element (RE), a subcarrier, a resource block (RB), or a subband.