CN-121986347-A - Federal split deep neural network computation

Abstract

Federal Machine Learning (FML) is combined with User Equipment (UE) split Deep Neural Network (DNN) computation. A plurality of UEs are assigned to the federal learning group for performing split execution of the neural network (106) such that the computing system executes a respective first portion (106A) of the neural network and each UE in the federal learning group executes a respective second portion (106B) of the neural network. An initial indication of a neural network configuration (114A-M) of a second portion of the neural network is sent to UEs in the federal learning group. Parameters of a respective second portion of the neural network are updated based on update information (126-M) received from the at least two UEs for the second portion of the neural network, and an update indication (130) of the updated parameters is sent to the UEs in the federal learning group.

Inventors

• WANG JIBING
• Eric Richard Stauffer

Assignees

• 谷歌有限责任公司

Dates

Publication Date

20260505

Application Date

20241001

Priority Date

20231003

Claims (20)

1. 1. A computer-implemented method performed by a computing system, the method comprising: Receiving (312) a request for a neural network configuration from a user equipment, UE; Assigning (340) the UEs to a federal learning group comprising a plurality of UEs performing split execution of a neural network based on the request, such that the computing system performs a respective first portion of the neural network for each UE in the federal learning group and each UE in the federal learning group performs a respective second portion of the neural network; -transmitting (314) an initial indication of a neural network configuration of the second portion of the neural network to the plurality of UEs in the federal learning group; Receiving respective update information for the respective second portions of the neural network executing at the plurality of UEs from at least two UEs of the plurality of UEs in the federal learning group; updating parameters of the second portion of the neural network based on the update information, and An update indication of updated parameters of the second portion of the neural network is issued to the plurality of UEs in the federal learning group.
2. 2. The method of claim 1, further comprising: Receiving a further request for a neural network configuration from a further UE, and The further UE is assigned to the federal learning group based on the further request for neural network configuration.
3. 3. The method of any one of claims 1 or 2, wherein the method further comprises: Executing a first portion of the neural network to generate an intermediate neural network output based on input data from the UE, and Causing the intermediary neural network to output a transmission to the UE.
4. 4. A method as claimed in any preceding claim, wherein the update information comprises one or more updated parameter values of the second part of the neural network and/or one or more changes in parameter values of the second part of the neural network.
5. 5. A method as claimed in any preceding claim, wherein the updated information comprises one or more gradients of a parameter of an objective function relative to the second part of the neural network.
6. 6. The method of claim 5, further comprising updating one or more parameters of the first portion of the neural network based on the one or more gradients of the objective function relative to parameters of the second portion of the neural network.
7. 7. The method of any preceding claim, wherein: the request for neural network configuration includes one or more local UE condition capability information of the UE, assistance information of the UE, and/or information indicative of local conditions of the UE, and Assigning the UE to a federal learning group includes determining that the local UE condition falls within a range of capabilities and/or local conditions associated with the federal learning group.
8. 8. The method of claim 7, wherein the one or more local UE conditions include one or more of UE capability information, UE assistance information, UE processing availability and/or capability information, DNN-specific processing availability and/or capability information, UE power information, and/or UE thermal information.
9. 9. The method of any preceding claim, wherein the initial indication of neural network configuration is the same for each UE in the federal learning group.
10. 10. The method of any one of claims 1 to 7, wherein: the respective second portions of the neural network performed by the plurality of UEs in the federal learning group have a set of one or more common layers across the federal learning group, and The proper subset of the UEs in the federal learning group have respective second portions of the neural networks that include one or more additional neural network layers that are not present in the set of one or more common layers.
11. 11. The method of any preceding claim, wherein the initial indication of the neural network configuration comprises an indication of one or more layer identities and/or types of the second portion of the neural network, a plurality of weights and/or bias values of nodes of the second portion of the neural network, an identity of a federal learning group to which the UE is assigned, a split location of the neural network, one or more triggers for sending the update information to the computing system, a type and/or format of the update information, and/or a type of local update procedure for generating the update information.
12. 12. The method of any preceding claim, wherein the computing system is a cellular network entity.
13. 13. The method of any preceding claim, wherein the respective update information is obtained from a local update procedure at each of a plurality of UEs in the federal learning group.
14. 14. A computer-implemented method performed by a user equipment, UE, the method comprising: Sending a request for neural network configuration to a network node; In response to the request, receiving an initial indication of a neural network configuration of a second portion of the neural network to be performed locally at the UE; Receiving, from the network node, an intermediate output of the neural network model generated by the first portion of the neural network; Processing the intermediate output by the second portion of the neural network model to generate a model output of the neural network; Generating an update to the second portion of the neural network based on the model output and using a local update process, and And sending the update to the network node.
15. 15. The method of claim 14, further comprising: Receiving one or more federal updates for the second portion of the neural network from the network node, and The received one or more federal updates are applied to the second portion of the neural network model.
16. 16. The method of any of claims 14 or 15, wherein the method further comprises, prior to receiving the intermediate output, sending input data of the first portion of the neural network to the network node.
17. 17. The method of any of claims 14 or 15, wherein the method further comprises, prior to receiving the intermediate output: processing the input data by a third portion of the neural network to generate an initial intermediate output, and The initial intermediate output is sent to the network node for input to the first portion of the neural network.
18. 18. The method of any of claims 14 to 17, wherein the request for neural network configuration comprises one or more local conditions of the UE.
19. 19. The method of claim 18, wherein the one or more local conditions of the UE include one or more of UE capability information, UE assistance information, UE processing availability and/or capability information, DNN-specific processing availability and/or capability information, UE power information, and/or UE thermal information.
20. 20. The method of any of claims 14 to 19, wherein the request for neural network configuration specifies a split location of the neural network.

Description

Federal split deep neural network computation Background For the fifth generation (5G) advanced and sixth generation (6G) cellular standards, the use of Deep Neural Networks (DNNs) is expected to play an important role. For example, to support augmented reality (XR) use cases such as Virtual Reality (VR) and Augmented Reality (AR), DNN may be used for image rendering. However, a User Equipment (UE) executing such applications may have capability constraints that affect the UE's execution of DNNs, such as constraints on processing power, available memory, available battery power, and so on. Such constraints may be temporary, such as available battery power, current memory or processor usage, etc., or permanent, such as hardware specifications of the UE. In addition, the UE allows the user to control access to the data. Disclosure of Invention In a first aspect, the present disclosure provides a computer-implemented method performed by a computing system, the method comprising receiving a request for a neural network configuration from a user equipment, UE, assigning the UE to a federal learning group comprising a plurality of UEs performing split execution of the neural network based on the request, such that the computing system performs a respective first portion of the neural network for each UE in the federal learning group and a respective second portion of the neural network, transmitting an initial indication of the neural network configuration of the second portion of the neural network to the plurality of UEs in the federal learning group, receiving respective update information for the respective second portion of the neural network performed at the plurality of UEs from at least two UEs in the federal learning group, updating parameters of the second portion of the neural network based on the update information, and issuing an update indication of the updated parameters of the second portion of the neural network to the plurality of UEs in the federal learning group. In a further aspect, the present description provides a computer-implemented method performed by a user equipment, UE, the method comprising sending a request for a neural network configuration to a network node, receiving an initial indication of the neural network configuration of a second portion of the neural network to be performed locally at the UE in response to the request, receiving an intermediate output of a neural network model generated by the first portion of the neural network from the network node, processing the intermediate output by the second portion of the neural network model to generate a model output of the neural network, generating an update to the second portion of the neural network based on the model output and using a local update procedure, and sending the update to the network node. Further aspects of the present disclosure provide systems, apparatuses, and computer-readable media for implementing any of the methods described herein. The systems, methods, and apparatus described herein may be implemented to realize one or more of the following advantages. Using a split Deep Neural Network (DNN) in conjunction with federal learning may ensure that during both DNN inference and DNN training, sensitive data portions, such as private user data, may be locally retained at the UE while allowing the UE to utilize network resources when performing DNN, e.g., in order to perform a larger and/or more complex DNN than is possible using only the local UE. Drawings FIG. 1A illustrates an overview of an example system and method for assigning UEs to federal learning groups; FIG. 1B illustrates an overview of an example system and method for performing split DNNs; FIG. 1C illustrates an overview of an example system and method for updating a split DNN using federal machine learning; FIG. 2 shows an overview of a further example system for performing splitting DNNs, where the UE-side splitting of DNNs is not the same for all UEs in a Federal learning group; FIG. 3 shows a signaling diagram for performing federal learning with split DNN architecture; Fig. 4A shows an example of a signaling diagram for performing user plane data at a two-part split DNN; fig. 4B shows a further example of a signaling diagram for performing user plane data at a three-part split DNN; Fig. 4C shows a further example of a signaling diagram for performing user plane data at a two-part split DNN; fig. 4D shows a further example of a signaling diagram for performing user plane data at a three-part split DNN; FIG. 5A illustrates an example of a federal machine learning process for two-part split DNN; FIG. 5B illustrates an example of a federal machine learning process for a three-part split DNN; FIG. 5C illustrates a further example of a federal machine learning process for two-part and/or three-part split DNNs; FIG. 6 illustrates a flow chart of an example network side method for performing federal learning using a split DNN architecture; FIG. 7