Optimizing training data for image classification

Patent metadata

Field	Value
Publication number	US-11954572-B2
Application number	US-202318198048-A
Country	US
Kind code	B2
Filing date	May 16, 2023
Priority date	Jun 6, 2019
Publication date	Apr 9, 2024
Grant date	Apr 9, 2024

Abstract

A method for machine learning-based classification may include training a machine learning model with a full training data set, the full training data set comprising a plurality of data points, to generate a first model state of the machine learning model, generating respective embeddings for the data points in the full training data set with the first model state of the machine learning model, applying a clustering algorithm to the respective embeddings to generate one or more clusters of the embeddings, identifying outlier embeddings from the one or more clusters of the embeddings, generating a reduced training data set comprising the full training data set less the data points associated with the outlier embeddings, training the machine learning model with the reduced training data set to a second model state, and applying the second model state to one or more data sets to classify the one or more data sets.

First claim

What is claimed is:

1. A method for machine learning-based classification, the method comprising:
generating, via a pre-trained machine learning model, respective embeddings for data points in a training data set used to pre-train the pre-trained machine learning model;
applying a clustering algorithm to the respective embeddings to generate a plurality of clusters of the embeddings;
identifying outlier embeddings from the plurality of clusters of the embeddings;
generating a reduced training data set comprising the full training data set less the data points associated with the outlier embeddings, the reduced training data set including the data points associated with embeddings in the plurality of clusters;
tuning the pre-trained machine learning model with the reduced training data set; and
applying the tuned pre-trained machine learning model to one or more data sets to classify the one or more data sets.

2. The method of claim 1 , wherein applying the tuned pre-trained machine learning model to classify one or more data sets comprises applying the tuned pre-trained machine learning model to classify one or more images.

3. The method of claim 1 , further comprising:
applying a distance learning algorithm to the respective embeddings to create a distanced embeddings set;
wherein applying a clustering algorithm to the respective embeddings comprises applying the clustering algorithm to the distanced embeddings set.

4. The method of claim 1 , wherein identifying outlier embeddings from the plurality of clusters of the embeddings comprises:
designating embeddings that are remote from a single cluster of embeddings as outlier embeddings; or
designating embeddings that are remote from all of the plurality of clusters as outlier embeddings.

5. The method of claim 1 , wherein identifying outlier embeddings from the plurality of clusters of the embeddings comprises:
determining a respective category associated with each of the embeddings;
determining a respective category associated with each cluster of embeddings; and
designating embeddings that are remote from a cluster of embeddings associated with the category with which the embeddings are associated as outlier embeddings.

6. The method of claim 1 , wherein identifying outlier embeddings from the plurality of clusters of the embeddings comprises:
identifying at least a predetermined percentage of embeddings as outlier embeddings;
identifying at least a predetermined quantity of embeddings as outlier embeddings; or
identifying embeddings that are a predetermined distance from one of the plurality of clusters as outlier embeddings.

7. The method of claim 1 , wherein tuning the pre-trained machine learning model with the reduced training data set comprises:
training, for a plurality of epochs, the pre-trained machine learning model with the reduced training data set;
storing a respective version of the pre-trained machine learning model for each epoch; and
determining the tuned pre-trained machine learning model as the respective version with a highest accuracy.

8. The method of claim 7 , wherein determining the tuned pre-training model as the respective version with the highest accuracy comprises:
retrieving an image and a classification label associated with the image;
generating, by each of the respective versions of the pre-trained machine learning model, a predicted classification label for the retrieved image; and
determining an accuracy of each of the respective versions as a difference between the retrieved classification label and the respective predicted classification label.

9. A system for machine learning-based classification, the system comprising:
a processor; and
a non-transitory, computer-readable memory storing instructions that, when executed by the processor, cause the processor to:
generate, via a machine learning model pre-trained in a first model state, respective embeddings for data points in a full training data set used to pre-train the machine learning model;
apply a clustering algorithm to the respective embeddings to generate a plurality of clusters of the embeddings;
identify outlier embeddings from the plurality of clusters of the embeddings;
generate a reduced training data set comprising the full training data set less the data points associated with the outlier embeddings, the reduced training data set including the data points associated with embeddings in the plurality of clusters;
train the machine learning model with the reduced training data set to a second model state; and
apply the second model state to one or more data sets to classify the one or more data sets.

10. The system of claim 9 , wherein applying the second model state to classify one or more data sets comprises applying the second model state to classify one or more images.

11. The system of claim 9 , wherein the memory stores further instructions that, when executed by the processor, cause the processor to:
apply a distance learning algorithm to the respective embeddings to create a distanced embeddings set;
wherein applying a clustering algorithm to the respective embeddings comprises applying the clustering algorithm to the distanced embeddings set.

12. The system of claim 9 , wherein identifying outlier embeddings from the plurality of clusters of the embeddings comprises:
designating embeddings that are remote from a single cluster of embeddings as outlier embeddings; or
designating embeddings that are remote from all of the plurality of clusters as outlier embeddings.

13. The system of claim 9 , wherein identifying outlier embeddings from the plurality of clusters of the embeddings comprises:
determining a respective category associated with each of the embeddings;
determining a respective category associated with each cluster of embeddings; and
designating embeddings that are remote from a cluster of embeddings associated with the category with which the embeddings are associated as outlier embeddings.

14. The system of claim 9 , wherein identifying outlier embeddings from the plurality of clusters of the embeddings comprises:
identifying at least a predetermined percentage of embeddings as outlier embeddings;
identifying at least a predetermined quantity of embeddings as outlier embeddings; or
identifying embeddings that are a predetermined distance from one of the plurality of clusters as outlier embeddings.

15. The system of claim 9 , wherein training the machine learning model with the reduced training data set comprises:
training, for a plurality of epochs, the machine learning model with the reduced training data set;
storing a respective version of the machine learning model for each epoch; and
determining the second model state as the respective version with a highest accuracy.

16. The system of claim 15 , wherein determining the second model state as the respective version with the highest accuracy comprises:
retrieving an image and a classification label associated with the image;
generating, by each of the respective versions of the machine learning model, a predicted classification label for the retrieved image; and
determining an accuracy of each of the respective versions as a difference between the retrieved classification label and the respective predicted classification label.

17. A machine learning-based method of classifying a plurality of images, the method comprising:
generating, via a machine learning model, respective embeddings for images in a full training data set used to pre-train the machine learning model;
applying

Assignees

• Home Depot Product Authority Llc

Inventors

• Hagen Matthew
• Afshar Estelle
• Qu Huiming
• Ayadi Ala Eddine
• Wang Jiaqi

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N20/20 Primary

  Ensemble learning · CPC title

• G06F18/2148

  characterised by the process organisation or structure, e.g. boosting cascade · CPC title

• G06F18/2155

  characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling · CPC title