Adaptively adjusting influence in federated learning model updates

Patent metadata

Field	Value
Publication number	US-11948096-B2
Application number	US-202016818537-A
Country	US
Kind code	B2
Filing date	Mar 13, 2020
Priority date	Mar 13, 2020
Publication date	Apr 2, 2024
Grant date	Apr 2, 2024

Abstract

Techniques for improved federated learning are provided. One or more queries are issued to a plurality of participants in a federated learning system, and one or more replies are received from the plurality of participants. A first aggregated model is generated based on the one or more relies and a first influence vector. Upon determining that a predefined criterion is satisfied, a second influence vector modifying a weight of a first participant of the plurality of participants is generated. A second aggregated model is generated based on the one or more replies and the second influence vector.

First claim

What is claimed is:

1. A method, comprising:
issuing, by a computing device, one or more queries to a plurality of participants in a federated learning system;
receiving a first set of replies from the plurality of participants;
generating a first aggregated model based on the first set of replies, a first influence vector, and an initial model; and
upon determining that a predefined criterion is satisfied, during a first epoch:
generating a second influence vector modifying a weight of a first participant of the plurality of participants,
generating a second aggregated model based on the first set of replies, the second influence vector, and the initial model,
generating a third influence vector modifying a weight of a first participant of the plurality of participants, and
generating a third aggregated model based on the first set of replies, the third influence vector, and an initial model; and

upon determining that a number of aggregated models in a set of aggregated models, including the first, second and third aggregated models, exceeds a predefined maximum value, pruning one or more of the first, second or third aggregated models during a second epoch based on model performance.

2. The method of claim 1 , wherein the first influence vector assigns equal weight to each of the plurality of participants, wherein the second influence vector assigns additional weight to the first participant, and wherein the third influence vector assigns less weight to the first participant.

3. The method of claim 1 , wherein determining that the predefined criterion is satisfied comprises identifying a change in one or more conditions of the first participant based on a reply from the first participant.

4. The method of claim 1 , wherein determining that the predefined criterion is satisfied comprises receiving, from a third party, new information about the first participant.

5. The method of claim 1 , wherein determining that the predefined criterion is satisfied comprises determining that a reply from the first participant is an outlier, as compared to replies from other participants in the plurality of participants, wherein determining that the reply from the first participant is an outlier comprises:
comparing the reply from the first participant to the replies from the other participants; and
determining that a difference between the reply from the first participant and the replies from the other participants exceeds a predefined threshold.

6. The method of claim 1 , wherein determining that the predefined criterion is satisfied comprises determining that a reply from the first participant is an outlier, as compared to replies from the first participant, wherein determining that the reply from the first participant is an outlier comprises:
comparing the reply from the first participant to a last reply from the first participant; and
determining that a difference between the reply from the first participant and the last reply from the first participant exceeds a predefined threshold.

7. The method of claim 1 , wherein evaluating each respective aggregated model comprises:
transmitting, to at least one participant of the plurality of participants, the set of aggregated models; and
requesting that the at least one participant evaluate the set of aggregated models using local data.

8. One or more computer-readable storage media collectively containing computer program code that, when executed by operation of one or more computer processors, performs an operation comprising:
issuing one or more queries to a plurality of participants in a federated learning system;
receiving a first set of replies from the plurality of participants;
generating a first aggregated model based on the first set of relies, a first influence vector, and an initial model; and
upon determining that a predefined criterion is satisfied, during a first epoch:
generating a second influence vector modifying a weight of a first participant of the plurality of participants,
generating a second aggregated model based on the first set of replies, the second influence vector, and the initial model,
generating a third influence vector modifying a weight of a first participant of the plurality of participants, and
generating a third aggregated model based on the first set of replies, the third influence vector, and an initial model; and

upon determining that a number of aggregated models in a set of aggregated models, including the first, second and third aggregated models, exceeds a predefined maximum value, pruning one or more of the first, second or third aggregated models during a second epoch based on model performance.

9. The computer-readable storage media of claim 8 , wherein the first influence vector assigns equal weight to each of the plurality of participants, wherein the second influence vector assigns additional weight to the first participant, and wherein the third influence vector assigns less weight to the first participant.

10. The computer-readable storage media of claim 8 , wherein determining that the predefined criterion is satisfied comprises determining that a reply from the first participant is an outlier, as compared to replies from other participants in the plurality of participants, wherein determining that the reply from the first participant is an outlier comprises:
comparing the reply from the first participant to the replies from the other participants; and
determining that a difference between the reply from the first participant and the replies from the other participants exceeds a predefined threshold.

11. The computer-readable storage media of claim 8 , wherein determining that the predefined criterion is satisfied comprises determining that a reply from the first participant is an outlier, as compared to replies from the first participant, wherein determining that the reply from the first participant is an outlier comprises:
comparing the reply from the first participant to a last reply from the first participant; and
determining that a difference between the reply from the first participant and the last reply from the first participant exceeds a predefined threshold.

12. The computer-readable storage media of claim 8 , wherein evaluating each respective aggregated model comprises:
transmitting, to at least one participant of the plurality of participants, the set of aggregated models; and
requesting that the at least one participant evaluate the set of aggregated models using local data.

13. A system comprising:
one or more computer processors; and
one or more memories collectively containing one or more programs which when executed by the one or more computer processors performs an operation, the operation comprising:
issuing one or more queries to a plurality of participants in a federated learning system;
receiving a first set of replies from the plurality of participants;
generating a first aggregated model based on the first set of relies, a first influence vector, and an initial model; and
upon determining that a predefined criterion is satisfied, during a first epoch:
generating a second influence vector modifying a weight of a first participant of the plurality of participants,
generating a second aggregated model based on the first set of replies, the second influence vector, and the initial model,
generating a third influence vector modifying a weight of a first participant of the plurality of participants, and
generating a third aggregated model based on the first set of replies, the third influence vector, and an initial model; and

upon determining that a number of aggregated models in a set of aggr

Assignees

• IBM

Inventors

• Zhou Yi
• Anwar Ali
• Baracaldo Angel Nathalie
• Ludwig Hekio H

Classifications

• G06N5/04 Primary

  Inference or reasoning models · CPC title

• G06N3/082

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N20/20 Primary

  Ensemble learning · CPC title