Qualcomm Incorporated

Eren Balevi

Taesang Yoo

Tao Luo

Hamed Pezeshki

Aziz Gholmieh

Junyi Li

Srinivas Yerramalli

Rajat Prakash

Gavin Bernard Horn

Xipeng Zhu

Shankar Krishnan

Xiaoxia Zhang

Rajeev Kumar

Vanitha Aravamudhan Kumar

Vishal Dalmiya

Jing X Sun

Tingfang Ji

Mahmoud Taherzadeh Boroujeni

Zhifei Fan

June Namgoong

Kirty Prabhakar Vedula

Aziz Gholmich

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.

Power control in over the air aggregation for federated learning

Disclosed are systems and techniques for wireless communications. For instance, a user equipment (UE) can perform federated learning to generate a first set of updated model parameters corresponding to a machine learning model. In some cases, the UE can receive a request for the first set of updated model parameters from a network entity, wherein the request includes a resource allocation associated with an uplink channel. In some examples, the UE can determine a signal phase corresponding to the uplink channel. In some aspects, the UE can transmit, based on the signal phase, the first set of updated model parameters using the resource allocation on the uplink channel.

Signal synchronization for over-the-air aggregation in a federated learning framework

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.

Gradient dropping for federated learning

Methods, systems, and devices for wireless communications are described. In some examples, a wireless communications system may support machine learning and may configure a user equipment (UE) for machine learning. The UE may transmit, to a base station, a request message that includes an indication of a machine learning model or a neural network function based at least in part on a trigger event. In response to the request message, the base station may transmit a machine learning model, a set of parameters corresponding to the machine learning model, or a configuration corresponding to a neural network function and may transmit an activation message to the UE to implement the machine learning model and the neural network function.

Configuring a user equipment for machine learning

Aspects presented herein may enable a network entity to configure a group of UEs to simultaneously transmit reference signals and to simultaneously transmit gradient vectors to the network entity, such that the network entity may receive the gradient vectors from the group of UEs as an aggregated gradient vector over the air. In one aspect, a base transmits, to a group of UEs, a configuration that configures the group of UEs to simultaneously transmit one or more group-common reference signals and to simultaneously transmit one or more gradient vectors associated with a federated learning procedure. The network entity receives, from the group of UEs, the one or more group-common reference signals and the one or more gradient vectors based on the configuration via multiple channels. The network entity calculates an average gradient vector based on the one or more group-common reference signals and the one or more gradient vectors.

Group-common reference signal for over-the-air aggregation in federated learning

Methods, systems, and devices for wireless communications are described. A user equipment (UE) may encode channel state feedback (CSF) information to compress the CSF information to a first encoding output associated with a first dimensional space, and apply entropy coding to the first encoding output of the channel state feedback information. The entropy coding may transform the first encoding output to a second encoding output associated with a second dimensional space that is smaller than the first dimensional space of the first encoding output. The UE may transmit a CSF message comprising the second encoding output. A network device may receive the CSF message and apply entropy decoding to the compressed CSF information to partially decompress the compressed CSF information to a first decoding output. The network device may decode the first decoding output to completely decompress the compressed CSF information to a second decoding output.

Techniques for encoding and decoding a channel between wireless communication devices

A method of wireless communication by a user equipment (UE) includes receiving, from a network device, a request to initiate gradient computations for a round of federated learning. The method also includes computing gradients in response to receiving the request to initiate gradient computation. The method further includes informing the network device of availability of the gradients. The method still further includes receiving, from the network device, information to enable transfer of the gradients to the network device. The method further includes transferring the gradients to the network device in response to receiving the information to enable transfer.

Method and apparatus for fetching gradient data for federated learning process

A method of wireless communication by a user equipment (UE) includes generating, by an upper analog media access control (MAC-A) layer of a protocol stack, a data packet with a header and a data field. The header indicates a neural network identifier (ID) and a request ID. The data field includes gradient data for a federated learning iteration. The method also includes transferring the data packet to lower layers of the protocol stack for transmission to a network device across a wireless network.

Upper analog media access control (MAC-A) layer functions for analog transmission protocol stack

A protocol stack architecture for processing machine learning (ML) data includes a ML layer to manage ML data communication with a network device. The ML layer is coupled to multiple ML training blocks, and ML and inference blocks for multiple neural networks, and an analog data communications stack coupled to the ML layer. The analog data communications stack has an upper media access control analog (MAC-A) layer coupled to the ML layer and configured to store data for each neural network, a lower MAC-A layer coupled to the upper MAC-A layer and configured to segment and reassemble analog ML data, and an analog physical layer coupled to the lower MAC-A layer and configured to communicate analog data with the network device. The architecture includes a digital data communications stack coupled to the ML layer and the lower MAC-A layer and configured to manage digital communications with the network device.

Protocol stack for analog communication in split architecture network for machine learning (ML) functions

Certain aspects of the present disclosure provide techniques and apparatus for determining neural network functions (NNFs) and configuring and using corresponding machine learning (ML) models for performing one or more ML-based wireless communications management procedures. An example method performed by a user equipment includes transmitting, to a base station (BS), UE capability information indicating at least one radio capability of the UE and at least one machine learning (ML) capability of the UE and receiving, from the BS based on the UE capability information, ML configuration information indicating at least one neural network function (NNF) and at least one ML model corresponding to the at least one NNF.

Machine learning model configuration in wireless networks

Methods, systems, and devices for wireless communications are described. A user equipment (UE) may communicate with a network entity within a wireless communications network. The UE may transmit a request for information to the network entity and, in response to the request, the UE may receive the requested information from the network entity. For example, the UE may request data from one or more data repositories associated with the network entity. In some examples, the information request may be associated with one or more measurements associated with operations of the network. In some instances, the UE may use a machine learning model to perform training or inference based on the information associated with the one or more measurements.

Network measurements for enhanced machine learning model training and inference

A UE may identify, in each round other than an initial round, a first plurality of local model update elements of a present round. The first plurality of local model update elements of the present round may be associated with an updated local machine learning model. The UE may transmit to a base station, in each round other than the initial round, over a multiple access channel via analog signaling, a second plurality of local model update elements of the present round based on a third plurality of local model update elements of the present round. The third plurality of local model update elements of the present round may correspond to a sum of the first plurality of local model update elements of the present round and a local model update error of a previous round immediately before the present round. The analog signaling may be associated with OTA aggregation.

Gradient accumulation for federated learning

A method of wireless communication by a user equipment (UE) includes receiving an uplink grant from a network device. The method also includes determining, by a lower analog media access control (MAC-A) layer of an analog data communications stack, a quantity of analog neural network gradients to transmit in a data packet using resources allocated by the uplink grant. The determining is based on an analog physical (PHY-A) layer encoding scheme and the resources allocated by the uplink grant. The method further includes segmenting the analog neural network gradients into the data packet. The method still further includes transferring the data packet to the PHY-A layer for transmission to the network device across a wireless network.

Lower analog media access control (MAC-A) layer and physical layer (PHY-A) functions for analog transmission protocol stack

Various aspects of the present disclosure generally relate to wireless communication. In some aspects, a user equipment (UE) may receive, from a base station, a federated learning configuration that configures a participation indication to be used by the UE to indicate a participation status of the UE associated with at least one federated learning round corresponding to a machine learning component. The UE may transmit the participation indication to the base station based at least in part on the federated learning configuration. Numerous other aspects are described.

Growing community of inventors

San Diego, CA, United States of America

Eren Balevi