US 2026/0141243 A1Published application

Communication-efficient training for wireless split-learning-based functions

AssigneeHuawei Technologies Co., Ltd.

InventorsOmar Ahmad Mohammad Alhussein, Mehdi Arashmid Akhavain Mohammadi

Life of this patent

Estimating the term…

Loading patent…

Abstract

Methods, apparatus, and systems for training a neural network that is split over elements of a communication network are disclosed. To train a split neural network, data needs to be iteratively transmitted between encoding and decoding parts of the neural network, which requires provisioning of resources from the communication network. Embodiments of the present disclosure involve determining whether to send model parameters between the encoding and decoding parts. Some embodiments may determine this based on the size of the neural network and the size of the dataset needed for training. Some embodiments may send the modelling parameters to a location intermediate between the encoding and decoding parts for processing.

Claims20

Dependent claims

A method, comprising:

determining a communication cost parameter (CCP) for training a neural network (NN), the NN having one or more NN encoders and an NN decoder, the one or more NN encoders making up an NN encoder series, each NN encoder having respective encoder model parameters, each NN encoder being deployed in a respective encoder network element of one or more encoder network elements, the NN decoder being deployed in a decoder network element, each encoder network element being coupled to the decoder network element across a respective first separation, each encoder network element being coupled to a respective proxy location of one or more proxy locations, each proxy location being coupled to the decoder network element over a respective second separation, and each second separation being shorter than the respective first separation;
and
when the CCP is greater than a threshold value, performing a set of actions for each NN encoder over the NN encoder series, the set of actions comprising:
sending, by the respective encoder network element, the respective encoder model parameters to the respective proxy location.

1 more element

A communication network, comprising:

one or more encoder network elements each having a neural network (NN) encoder, each NN encoder having encoder model parameters and being configured to generate one or more latent codes from input data in accordance with the respective encoder model parameters, each encoder network element being configured to:
obtain input data;
transmit latent codes;
and
transmit, when a communication cost parameter (CCP) is greater than a threshold value, encoder model parameters;
one or more proxy locations each being coupled to one or more encoder network elements, each proxy location being configured to:
receive encoder model parameters from the respective one or more encoder network elements;
and
a decoder network element having an NN decoder, the decoder network element being coupled to each encoder network element and each proxy location, the NN decoder having decoder model parameters and being configured to generate one or more prediction labels from latent codes in accordance with the decoder model parameters, the decoder network element configured to:
receive latent codes from each encoder network element.

7 more elements

Dependencies are inferred from the claim text.