Personalized Federated Learning for Assistant Systems

Patent metadata

Field	Value
Publication number	US-2021117780-A1
Application number	US-202016815990-A
Country	US
Kind code	A1
Filing date	Mar 11, 2020
Priority date	Oct 18, 2019
Publication date	Apr 22, 2021
Grant date	—

Abstract

In one embodiment, a method includes receiving, by a first client system, from one or more remote servers, a current version of a global neural network model including multiple federated model parameters, accessing, from a local data store, multiple examples and a local personalization model including multiple local model parameters, wherein each of the examples includes one or more features and one or more labels, training the global neural network model and the local personalization model together on the examples to generate multiple updated federated model parameters and multiple updated local model parameters, storing, in the local data store, the trained local personalization model including the updated local model parameters, and sending, to one or more of the remote servers, the trained global neural network model including the updated federated model parameters.

First claim

What is claimed is:

1 . A method comprising, by a first client system:
receiving, from one or more remote servers, a current version of a global neural network model comprising a plurality of federated model parameters; accessing, from a local data store, a plurality of examples and a local personalization model comprising a plurality of local model parameters, wherein each of the plurality of examples comprises one or more features and one or more labels; training the global neural network model and the local personalization model together on the plurality of examples to generate a plurality of updated federated model parameters and a plurality of updated local model parameters; storing, in the local data store, the trained local personalization model comprising the plurality of updated local model parameters; and sending, to one or more of the remote servers, the trained global neural network model comprising the plurality of updated federated model parameters.

2 . The method of claim 1 , wherein the global neural network model is configured to generate, responsive to a first example being input into the global neural network model, one or more candidate labels corresponding to the first example.

3 . The method of claim 1 , wherein the global neural network model is a natural-language generation model, and wherein, for each of the plurality of examples, at least one of the labels is a linguistic response comprising one or more n-grams.

4 . The method of claim 1 , wherein the global neural network model is a data classification model, and wherein, for each of the plurality of examples, at least one of the labels is a data classification associated with the example.

5 . The method of claim 1 , wherein the global neural network model and the local personalization model are iteratively trained together using Stochastic Gradient Descent (SGD).

6 . The method of claim 1 , wherein training the global neural network model and the local personalized model together on the plurality of examples comprises:
inputting each of the plurality of examples into the global neural network model; generating, for each of the plurality of examples, one or more candidate labels based on the input example, one or more of the local model parameters, and one or more of the federated model parameters; and generating the plurality of updated federated model parameters and the plurality of updated local model parameters based on the generated candidate labels.

7 . The method of claim 1 , wherein the plurality of updated federated model parameters and the plurality of updated local model parameters are generated based on, for each of the plurality of examples, a determination of whether one or more of the candidate labels generated by the global neural network model matches one or more of the labels.

8 . The method of claim 1 , wherein the plurality of updated federated model parameters and the plurality of updated local model parameters are generated based on, for each of the plurality of examples, a measure of error between one or more of the candidate labels generated by the global neural network model and one or more of the labels.

9 . The method of claim 1 , wherein one or more of the local model parameters correspond to one or more respective user representations in a user representation matrix.

10 . The method of claim 1 , wherein training the global neural network model and the local personalization model together on the plurality of examples is based at least in part on one or more respective user characteristics associated with the first client system.

11 . The method of claim 10 , wherein the user characteristics associated with the first client system comprise one or more of linguistic fluency, location, age, gender, occupation, income, marital status, number of children, or number of relationships.

12 . The method of claim 1 , wherein one or more of the local model parameters correspond to one or more respective linguistic characteristics associated with the first client system.

13 . The method of claim 12 , wherein the linguistic characteristics associated with the first client system comprise one or more of linguistic formality, punctuation usage, emoticon usage, or grammatical style.

14 . The method of claim 1 , wherein one or more of the local model parameters correspond to a probability that an example stored on the first client system is a linguistic conversation having greater than a threshold number of participants.

15 . The method of claim 1 , wherein the local personalization model further comprises a personalized portion of the global neural network model.

16 . The method of claim 1 , wherein the plurality of examples comprises, for each of a plurality of task categories, a set of examples associated with the task category, and wherein the training of the global neural network model and the local personalization model is repeated for each set of examples.

17 . The method of claim 1 , wherein the updated local model parameters comprise one or more new local model parameters created by the trained global neural network model.

18 . The method of claim 17 , wherein the trained global neural network model sent to the one or more remote servers is configured to modify one or more local personalization models of one or more other client systems, respectively, to include the one or more new local model parameters.

19 . A system comprising: one or more processors; and a memory coupled to the processors comprising instructions executable by the processors, the processors operable when executing the instructions to:
receive, from one or more remote servers, a current version of a global neural network model comprising a plurality of federated model parameters; access, from a local data store, a plurality of examples and a local personalization model comprising a plurality of local model parameters, wherein each of the plurality of examples comprises one or more features and one or more labels; train the global neural network model and the local personalization model together on the plurality of examples to generate a plurality of updated federated model parameters and a plurality of updated local model parameters; store, in the local data store, the trained local personalization model comprising the plurality of updated local model parameters; and send, to one or more of the remote servers, the trained global neural network model comprising the plurality of updated federated model parameters.

20 . One or more computer-readable non-transitory storage media embodying software that is operable when executed to:
receive, from one or more remote servers, a current version of a global neural network model comprising a plurality of federated model parameters; access, from a local data store, a plurality of examples and a local personalization model comprising a plurality of local model parameters, wherein each of the plurality of examples comprises one or more features and one or more labels; train the global neural network model and the local personalization model together on the plurality of examples to generate a plurality of updated federated model parameters and a plurality of updated local model parameters; store, in the local data store, the trained local personalization model comprising the plurality of updated local model parameters; and send, to one or more of the remote servers, the trained global neural network model comprising the plurality of updated federated model parameters.

Assignees

• Facebook Tech Llc

Inventors

• Malik Kshitiz
• Moon Seungwhan
• Liu Honglei
• Kumar Anuj
• Zhan Hongyuan
• Aly Ahmed

Classifications

• G06Q10/40 Primary

  Business processes related to social networking or social networking services · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06V20/41

  Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items (segmenting video sequences G06V20/49) · CPC title

• G06V40/174

  Facial expression recognition · CPC title