Methods, apparatus and machine-readable media relating to machine-learning in a communication network
A method performed by a first network entity in a communications network includes training a model to obtain a local model update including an update to values of one or more parameters of the model, in which training the model includes inputting training data into a machine learning algorithm. The method further includes applying a serialisation function to the local model update to construct a serial representation of the local model update, thereby removing information indicative of a structure of the model, and transmitting the serial representation of the local model update to an aggregator entity in the communications network.
1 . A method performed by a first network entity in a communications network, the method comprising:
training a model to obtain a local model update comprising an update to values of one or more parameters of the model, wherein training the model comprises inputting training data into a machine learning algorithm;
generating a serial representation of the local model update without information indicative of a structure of the model by applying a serialization function to the local model update; and
transmitting the serial representation of the local model update to an aggregator entity in the communications network.
2 . A first network entity for a communications network, the first network entity comprising:
processing circuitry; and
a memory coupled to the processing circuitry and having instructions stored therein that are executable by the processing circuitry to cause the first network entity to perform operations comprising:
training a model to obtain a local model update comprising an update to values of one or more parameters of the model, wherein training the model comprises inputting training data into a machine learning algorithm;
generating a serial representation of the local model update without information indicative of a structure of the model by applying a serialization function to the local model update; and
transmitting the serial representation of the local model update to an aggregator entity in the communications network.
3 . The first network entity of claim 2 , the operations further comprising:
receiving, from the aggregator entity, a combined model update, wherein the combined model update is based on the local model update and at least one additional local model update obtained by the aggregator entity from at least one second network entity in the communications network.
4 . The first network entity of claim 3 , wherein receiving the combined model update comprises receiving, from the aggregator entity, a serial representation of the combined model update.
5 . The first network entity of claim 4 , the operations further comprising:
obtaining a second update to the values of the one or more parameters of the model by applying an inverse of the serialization function to the serial representation of the combined model update.
6 . The first network entity of claim 3 , wherein the combined model update comprises at least one of:
differential values between an initial version of the model and an updated version of the model; and
values for an updated version of the model.
7 . The first network entity of claim 2 , wherein the structure of the model comprises one or more first data structures having respective dimensions,
wherein the serial representation comprises one or more second data structures having respective dimensions, and
wherein dimensions of the second data structures are different to dimensions of the first data structures.
8 . The first network entity of claim 7 , wherein:
the structure of the model comprises a first number of first data structures;
the serial representation comprises a second number of second data structures that is different than the first number of first data structures;
each of the first data structures corresponds to a layer of a neural network; and/or
the first and second data structures comprise matrices.
9 . The first network entity of claim 2 , wherein the local model update comprises at least one of:
differential values between an initial version of the model and a trained version of the model; and
values for a trained version of the model.
10 . The first network entity of claim 2 , wherein the aggregating entity is implemented in a Network Data Analytics Function (“NWDAF”).
11 . A system in a communications network, the system comprising:
an aggregator entity; and
a plurality of network entities configured to perform operations comprising:
training a model to obtain a local model update comprising an update to values of one or more parameters of the model, wherein training the model comprises inputting training data into a machine learning algorithm;
generating a serial representation of the local model update without information indicative of a structure of the model by applying a serialization function to the local model update; and
transmitting the serial representation of the local model update to the aggregator entity, wherein the aggregator entity is configured to combine the local model updates received from the plurality of network entities to obtain a combined model update.
12 . The system of claim 11 , the operations further comprising:
receiving, from the aggregator entity, a combined model update, wherein the combined model update is based on the local model update and at least one additional local model update obtained by the aggregator entity from at least one second network entity in the communications network,
wherein receiving the combined model update comprises receiving, from the aggregator entity, a serial representation of the combined model update.
13 . The method of claim 1 , further comprising:
receiving, from the aggregator entity, a combined model update, wherein the combined model update is based on the local model update and at least one additional local model update obtained by the aggregator entity from at least one second network entity in the communications network.
14 . The method of claim 13 , wherein receiving the combined model update comprises receiving, from the aggregator entity, a serial representation of the combined model update.
15 . The method of claim 14 , further comprising:
obtaining a second update to the values of the one or more parameters of the model by applying an inverse of the serialization function to the serial representation of the combined model update.
16 . The method of claim 13 , wherein the combined model update comprises at least one of:
differential values between an initial version of the model and an updated version of the model; and
values for an updated version of the model.
17 . The method of claim 1 , wherein the structure of the model comprises one or more first data structures having respective dimensions,
wherein the serial representation comprises one or more second data structures having respective dimensions, and
wherein dimensions of the second data structures are different to dimensions of the first data structures.
18 . The method of claim 17 , wherein:
the structure of the model comprises a first number of first data structures;
the serial representation comprises a second number of second data structures that is different than the first number of first data structures;
each of the first data structures corresponds to a layer of a neural network; and/or
the first and second data structures comprise matrices.
19 . The method of claim 1 , wherein the local model update comprises at least one of:
differential values between an initial version of the model and a trained version of the model; and
values for a trained version of the model.
20 . The method of claim 1 , wherein the aggregating entity is implemented in a Network Data Analytics Function (“NWDAF”).