Regularization relaxation scheme

• US 10,438,129 B1
• Filed: 12/30/2014
• Issued: 10/08/2019
• Est. Priority Date: 12/30/2013
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google LLC (Alphabet Inc.)
• Inventors: Singer, Yoram, Shaked, Tal, Chandra, Tushar Deepak, Ie, Tze Way Eugene
• Primary Examiner(s): Rifkin, Ben M
• Application Number:
  US14/586,043
• Time in Patent Office: 1,743 Days
• Field of Search: 706 12
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training machine learning systems. One of the methods includes receiving a plurality of training examples; and training a machine learning system on each of the plurality of training examples to determine trained values for weights of a machine learning model, wherein training the machine learning system comprises: assigning an initial value for a regularization penalty for a particular weight for a particular feature; and adjusting the initial value for the regularization penalty for the particular weight for the particular feature during the training of the machine learning system.

Claims

1. A method for training a machine learning model, the method comprising:

receiving, by a machine learning system, a plurality of training examples, wherein each training example comprises one or more features that identify properties of a respective training instance;

training, by the machine learning system, the machine learning model on the plurality of training examples to determine trained values for weights of the machine learning model, wherein the machine learning model comprises a respective weight for each feature found in any of the training examples processed by the machine learning model, wherein the machine learning model is configured to process each of the training examples to determine a predicted output for the training example from current values of the weights for the features in the training example, and wherein training the machine learning model comprises;

assigning a respective initial value for a regularization penalty for a particular weight for a particular feature, wherein the machine learning model is configured to determine the predicted output for each training example that comprises the particular feature by performing computations that include combining a current value for the regularization penalty for the particular weight for the particular feature with the current value of the particular weight for the particular feature,

determining whether a frequency of training examples processed by the machine learning model that comprise the particular feature exceeds a threshold frequency, and

responsive to determining that the frequency of training examples processed during the training of the machine learning model that comprise the particular feature exceeds the threshold frequency, decreasing the initial value for the regularization penalty for the particular weight for the particular feature during the training of the machine learning model; and

generating, after completion of the training of the machine learning model, a trained machine learning model based on the trained values for the weights whose values have been adjusted based on the respective regularization penalties.

2. The method of claim 1, wherein the training examples are each labeled with either a positive label or negative label.

3. The method of claim 2, wherein determining whether the frequency of training examples processed by the machine learning model that comprise the particular feature exceeds the threshold frequency comprises:

determining whether more than a threshold proportion of training examples processed by the machine learning model that have a positive label comprise the particular feature.

4. The method of claim 1, further comprising:

determining whether a frequency of training examples processed by the machine learning model within a predetermined amount of time that comprise the particular feature does not exceed a threshold frequency; and

responsive to determining that the frequency of training examples processed by the machine learning model within the predetermined amount of time that comprise the particular features does not exceed the threshold frequency, increasing the initial value of the regularization penalty for the weight for the particular feature.

5. The method of claim 4, wherein determining whether a frequency of training examples processed by the machine learning model within a predetermined amount of time that comprise the particular feature does not exceed a threshold frequency comprises:

determining whether a number of occurrences of training examples processed by the machine learning model within the predetermined amount of time that comprise the particular feature does not exceed a threshold frequency.

6. The method of claim 1, wherein the particular feature belongs to a base template, wherein a base template is a feature category that includes features of a single feature type.

7. The method of claim 1, wherein the particular feature belongs to a cross-template, wherein the cross-template is a feature category that includes a plurality of features, with each of the plurality of features being of a respective feature type.

8. The method of claim 1, wherein the training instances are potential video viewings, and wherein the predicted output is a prediction of whether or not a particular user will view a particular video.

9. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations:

receiving a plurality of training examples, wherein each training example comprises one or more features that identify properties of a respective training instance;

training a machine learning model on each of the plurality of training examples to determine trained values for weights of the machine learning model, wherein the machine learning model comprises a respective weight for each feature found in any of the training examples processed by the machine learning model, wherein the machine learning model is configured to process each of the training examples to determine a predicted output for the training example from current values of the weights for the features in the training example, and wherein training the machine learning model comprises;

assigning a respective initial value for a regularization penalty for a particular weight for a particular feature, wherein the machine learning model is configured to determine the predicted output for each training example that comprises the particular feature by performing computations that include combining a current value for the regularization penalty for the particular weight for the particular feature with the current value of the particular weight for the particular feature,

determining whether a frequency of training examples processed by the machine learning model that comprise the particular feature exceeds a threshold frequency, and

responsive to determining that the frequency of training examples processed during the training of the machine learning model that comprise the particular feature exceeds the threshold frequency, decreasing the initial value for the regularization penalty for the particular weight for the particular feature during the training of the machine learning model; and

generating, after completion of the training of the machine learning model, a trained machine learning model based on the trained values for the weights whose values have been adjusted based on the respective regularization penalties.

10. The system of claim 9, wherein the training examples are each labeled with either a positive label or negative label.

11. The system of claim 10, wherein determining whether the frequency of training examples processed by the machine learning model that comprise the particular feature exceeds the threshold frequency comprises:

determining whether more than a threshold proportion of training examples processed by the machine learning model that have a positive label comprise the particular feature.

12. The system of claim 9, further comprising:

determining whether a frequency of training examples processed by the machine learning model within a predetermined amount of time that comprise the particular feature does not exceed a threshold frequency; and

responsive to determining that the frequency of training examples processed by the machine learning model within the predetermined amount of time that comprise the particular features does not exceed the threshold frequency, increasing the initial value of the regularization penalty for the weight for the particular feature.

13. The system of claim 12, wherein determining whether a frequency of training examples processed by the machine learning model within a predetermined amount of time that comprise the particular feature does not exceed a threshold frequency comprises:

determining whether a number of occurrences of training examples processed by the machine learning model within the predetermined amount of time that comprise the particular feature does not exceed a threshold frequency.

14. The system of claim 9, wherein the particular feature belongs to a base template, wherein a base template is a feature category that includes features of a single feature type.

15. The system of claim 9, wherein the particular feature belongs to a cross-template, wherein the cross-template is a feature category that includes a plurality of features, with each of the plurality of features being of a respective feature type.

16. The system of claim 9, wherein the training instances are potential video viewings, and wherein the predicted output is a prediction of whether or not a particular user will view a particular video.

17. A non-transitory computer storage medium encoded with a computer program, the computer program comprising instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising:

receiving a plurality of training examples, wherein each training example comprises one or more features that identify properties of a respective training instance;

training a machine learning model on the plurality of training examples to determine trained values for weights of the machine learning model, wherein the machine learning model comprises a respective weight for each feature found in any of the training examples processed by the machine learning model, wherein the machine learning model is configured to process each of the training examples to determine a predicted output for the training example from current values of the weights for the features in the training example, and wherein training the machine learning model comprises;

assigning a respective initial value for a regularization penalty for a particular weight for a particular feature, wherein the machine learning model is configured to determine the predicted output for each training example that comprises the particular feature by performing computations that include combining a current value for the regularization penalty for the particular weight for the particular feature with the current value of the particular weight for the particular feature;

determining whether a frequency of training examples processed by the machine learning model that comprise the particular feature exceeds a threshold frequency;

responsive to determining that the frequency of training examples processed during the training of the machine learning model that comprise the particular feature exceeds the threshold frequency, decreasing the initial value for the regularization penalty for the particular weight for the particular feature during the training of the machine learning model; and

generating, after completion of the training of the machine learning model, a trained machine learning model based on trained values for the weights whose values have been adjusted based on the respective.