Method, device, and program product for training model

Patent metadata

Field	Value
Publication number	US-12547906-B2
Application number	US-202217687918-A
Country	US
Kind code	B2
Filing date	Mar 7, 2022
Priority date	Jan 20, 2022
Publication date	Feb 10, 2026
Grant date	Feb 10, 2026

Abstract

The present disclosure relates to a method, a device, and a program product for training a model. The method includes: receiving at least one unlabeled sample and at least one labeled sample for training a pre-training model, the pre-training model being used to extract features of the samples; creating an undirected graph associated with the pre-training model using the at least one unlabeled sample and a set of training samples associated with the pre-training model; dividing the undirected graph to form a plurality of sub-graphs based on corresponding features of the unlabeled sample and the set of training samples, the plurality of sub-graphs corresponding to a plurality of classifications of the samples, respectively; and training, based on the plurality of sub-graphs and the at least one labeled sample, the pre-training model to generate a training model. A corresponding device and a corresponding computer program product are provided.

First claim

What is claimed is:

1 . A method for training a model, comprising:
receiving, in a processor-based machine learning system, at least one unlabeled sample and at least one labeled sample for training a pre-training model of the processor-based machine learning system, the pre-training model being used to extract features of the samples; creating, in the processor-based machine learning system, an undirected graph associated with the pre-training model using the at least one unlabeled sample and a set of training samples associated with the pre-training model, wherein nodes in the undirected graph correspond to respective samples of the at least one unlabeled sample and the set of training samples, and edges between respective pairs of the nodes comprise respective weights each indicating a level of similarity between a corresponding one of the pairs of nodes; dividing, in the processor-based machine learning system, the undirected graph to form a plurality of sub-graphs based on corresponding features of the unlabeled sample and the set of training samples, the plurality of sub-graphs corresponding to a plurality of classifications of the samples, respectively, wherein the plurality of sub-graphs comprise at least a positive sub-graph and a negative sub-graph, each of the positive sub-graph and the negative sub-graph comprising a respective different subset of the nodes corresponding to the respective samples of the at least one unlabeled sample and the set of training samples, the positive and negative sub-graphs being determined utilizing a contrastive learning loss function that separates positive sample pairs each classified as representing same objects from negative sample pairs each classified as representing different objects; training, in the processor-based machine learning system, based on the plurality of sub-graphs and the at least one labeled sample, the pre-training model to generate a training model of the processor-based machine learning system; wherein the training model is utilized to process at least portions of a plurality of images from captured video data as multiple additional samples, the processing comprising detecting at least first and second objects associated with respective regions of interest in at least one of the images and tracking the at least first and second detected objects across multiple additional images of the plurality of images, using the training model and one or more instances of dividing a corresponding undirected graph into sub-graphs each comprising a positive sub-graph and a negative sub-graph.

2 . The method according to claim 1 , wherein creating the undirected graph comprises: determining, for the unlabeled sample and a first sample and a second sample in the set of training samples, a first node and a second node in the undirected graph corresponding to the first sample and the second sample, respectively.

3 . The method according to claim 2 , wherein creating the undirected graph further comprises:
determining, based on the pre-training model, a first feature and a second feature of the first sample and the second sample, respectively; and determining an edge between the first node and the second node based on a similarity between the first feature and the second feature.

4 . The method according to claim 3 , wherein determining the edge comprises:
updating the edge based on a perturbation factor describing sample noise.

5 . The method according to claim 3 , wherein dividing the undirected graph to form the plurality of sub-graphs comprises:
determining, based on edges of the undirected graph, an entropy loss associated with the undirected graph; and dividing, based on the entropy loss, the undirected graph to form the plurality of sub-graphs.

6 . The method according to claim 5 , wherein determining the entropy loss comprises:
determining an indicator vector associated with a sub-graph in the plurality of sub-graphs; and determining the entropy loss based on a separation degree between the indicator vector and the plurality of sub-graphs.

7 . The method according to claim 1 , wherein training the pre-training model to generate the training model comprises:
determining, based on the plurality of sub-graphs, an entropy loss for updating the pre-training model; determining, based on a priori knowledge, a priori knowledge loss for updating the pre-training model; and updating the pre-training model based on the entropy loss and the a priori knowledge loss.

8 . The method according to claim 7 , wherein updating the pre-training model comprises:
determining, based on the entropy loss and the a priori knowledge loss, an overall loss for updating the pre-training model; and updating a parameter of the pre-training model using the overall loss.

9 . The method according to claim 1 , further comprising: iteratively executing the method based on the at least one unlabeled sample and the at least one labeled sample.

10 . The method according to claim 1 , further comprising: extracting, in response to receiving a new sample to be processed, features of the new sample using the training model.

11 . An electronic device, comprising:
at least one processor; and memory coupled to the at least one processor and having instructions stored therein, wherein the instructions, when executed by the at least one processor, cause the electronic device to perform actions comprising: receiving, in a processor-based machine learning system, at least one unlabeled sample and at least one labeled sample for training a pre-training model of the processor-based machine learning system, the pre-training model being used to extract features of the samples; creating, in the processor-based machine learning system, an undirected graph associated with the pre-training model using the at least one unlabeled sample and a set of training samples associated with the pre-training model, wherein nodes in the undirected graph correspond to respective samples of the at least one unlabeled sample and the set of training samples, and edges between respective pairs of the nodes comprise respective weights each indicating a level of similarity between a corresponding one of the pairs of nodes; dividing, in the processor-based machine learning system, the undirected graph to form a plurality of sub-graphs based on corresponding features of the unlabeled sample and the set of training samples, the plurality of sub-graphs corresponding to a plurality of classifications of the samples, respectively, wherein the plurality of sub-graphs comprise at least a positive sub-graph and a negative sub-graph, each of the positive sub-graph and the negative sub-graph comprising a respective different subset of the nodes corresponding to the respective samples of the at least one unlabeled sample and the set of training samples, the positive and negative sub-graphs being determined utilizing a contrastive learning loss function that separates positive sample pairs each classified as representing same objects from negative sample pairs each classified as representing different objects; training, in the processor-based machine learning system, based on the plurality of sub-graphs and the at least one labeled sample, the pre-training model to generate a training model of the processor-based machine learning system; wherein the training model is utilized to process at least portions of a plurality of images from captured video data as multiple additional samples, the processing comprising detecting at least first and second objects associated with respective regions of interest in at least one of the images and tracking the at least first and second detected objects acros

Assignees

• Dell Products Lp

Inventors

• Yang Wenbin
• Wang Zijia
• Ni Jiacheng
• Chen Qiang
• Jia Zhen

Classifications

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N20/00

  Machine learning · CPC title

• G06N5/022 Primary

  Knowledge engineering; Knowledge acquisition · CPC title

• G06N3/088 Primary

  Non-supervised learning, e.g. competitive learning · CPC title