Sampling from a generator neural network using a discriminator neural network

Patent metadata

Field	Value
Publication number	US-11514313-B2
Application number	US-201916580649-A
Country	US
Kind code	B2
Filing date	Sep 24, 2019
Priority date	Sep 27, 2018
Publication date	Nov 29, 2022
Grant date	Nov 29, 2022

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for providing a data sample in response to a request for a data sample. In one aspect, a method comprises: receiving a request for a new data sample; until a candidate new data sample is generated that satisfies an acceptance criterion, performing operations comprising: generating a candidate new data sample using a generator neural network; processing the candidate new data sample using a discriminator neural network to generate an imitation score; and determining, from the imitation score, whether the candidate new data sample satisfies the acceptance criterion; and providing the candidate new data sample that satisfies the acceptance criterion in response to the received request.

First claim

What is claimed is:

1. A method performed by one or more data processing apparatus, the method comprising:
receiving a request for a data sample;
until a new data sample is generated that satisfies an acceptance criterion, performing operations comprising:
generating a candidate new data sample using a generator neural network;
processing the candidate new data sample using a discriminator neural network to generate an imitation score,
wherein the discriminator neural network has been trained to generate an imitation score for a data sample that characterizes a likelihood for whether the data sample was: (i) drawn from a training set of given data samples, or (ii) generated by the generator neural network; and

determining, from the imitation score, whether the candidate new data sample satisfies the acceptance criterion; and

providing the new data sample that satisfies the acceptance criterion in response to the received request.

2. The method of claim 1 , wherein candidate new data samples generated by the generator neural network are images.

3. The method of claim 1 , wherein determining, from the imitation score, whether the candidate new sample satisfies the acceptance criterion comprises:
determining an acceptance score from the imitation score;
obtaining a control sample by sampling from a probability distribution over a range of numerical values; and
determining the candidate new data sample satisfies the acceptance criterion if the acceptance score has a higher value than the control sample.

4. The method of claim 3 , wherein determining an acceptance score from the imitation score comprises:
updating a value of a bounding variable by determining a current value of the bounding variable to be a maximum of: (i) a previous value of the bounding variable, and (ii) the imitation score; and
determining the acceptance score from the imitation score and the updated bounding variable.

5. The method of claim 4 , wherein determining the acceptance score from the imitation score and the updated bounding variable comprises:
determining the acceptance score to be σ(a), wherein σ(·) is a sigmoid function, wherein a is defined as:
a:=D−D M −log(1−exp( D−D M )),

wherein D is the imitation score, and wherein D M is the current value of the bounding variable.

6. The method of claim 4 , wherein determining the acceptance score from the imitation score and the updated bounding variable comprises:
determining the acceptance score to be σ(a−γ), wherein σ(·) is a sigmoid function, wherein γ is a numerical value, wherein a is defined as:
a:=D−D M −log(1−exp( D−D M −∈)),

wherein D is the imitation score, wherein D M is the current value of the bounding variable, and wherein ∈ is a constant added for numerical stability.

7. The method of claim 6 , further comprising:
determining the numerical value γ to be a predetermined percentile of previously determined acceptance scores.

8. The method of claim 3 , wherein obtaining a control sample by sampling from a probability distribution over a range of numerical values comprises:
obtaining the control sample by sampling from a uniform probability distribution over the range [0,1].

9. The method of claim 1 , wherein the discriminator neural network has been trained using a logistic loss function.

10. The method of claim 1 , wherein processing the imitation score using a sigmoid function defines a probability that the data sample was drawn from the training set of given data samples.

11. 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 comprising:
receiving a request for a data sample;
until a new data sample is generated that satisfies an acceptance criterion, performing operations comprising:
generating a candidate new data sample using a generator neural network;
processing the candidate new data sample using a discriminator neural network to generate an imitation score,
wherein the discriminator neural network has been trained to generate an imitation score for a data sample that characterizes a likelihood for whether the data sample was: (i) drawn from a training set of given data samples, or (ii) generated by the generator neural network; and

determining, from the imitation score, whether the candidate new data sample satisfies the acceptance criterion; and

providing the new data sample that satisfies the acceptance criterion in response to the received request.

12. The system of claim 11 , wherein candidate new data samples generated by the generator neural network are images.

13. The system of claim 11 , wherein determining, from the imitation score, whether the candidate new sample satisfies the acceptance criterion comprises:
determining an acceptance score from the imitation score;
obtaining a control sample by sampling from a probability distribution over a range of numerical values; and
determining the candidate new data sample satisfies the acceptance criterion if the acceptance score has a higher value than the control sample.

14. The system of claim 13 , wherein determining an acceptance score from the imitation score comprises:
updating a value of a bounding variable by determining a current value of the bounding variable to be a maximum of: (i) a previous value of the bounding variable, and (ii) the imitation score; and
determining the acceptance score from the imitation score and the updated bounding variable.

15. The system of claim 14 , wherein determining the acceptance score from the imitation score and the updated bounding variable comprises:
determining the acceptance score to be σ(a), wherein σ(·) is a sigmoid function, wherein is defined as:
a:=D−D M −log(1−exp( D−D M )),

wherein D is the imitation score, and wherein D M is the current value of the bounding variable.

16. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
receiving a request for a data sample;
until a new data sample is generated that satisfies an acceptance criterion, performing operations comprising:
generating a candidate new data sample using a generator neural network;
processing the candidate new data sample using a discriminator neural network to generate an imitation score,
wherein the discriminator neural network has been trained to generate an imitation score for a data sample that characterizes a likelihood for whether the data sample was: (i) drawn from a training set of given data samples, or (ii) generated by the generator neural network; and

determining, from the imitation score, whether the candidate new data sample satisfies the acceptance criterion; and

providing the new data sample that satisfies the acceptance criterion in response to the received request.

17. The non-transitory computer storage media of claim 16 , wherein candidate new data samples generated by the generator neural network are images.

18. The non-transitory computer storage media of claim 16 , wherein determining, from the imitation score, whether the candidate new sample satisfies the acceptance criterion comprises:
determining an acceptance score from the imitation score;
obtaining a control sample by sampling from a probability distribution over a range of numerical values; and
determining the candidate new data sample satisfies the acceptance criterion

Assignees

• Google Llc

Inventors

• Azadi Samaneh
• Goodfellow Ian
• Olsson Catherine
• Odena Augustus Quadrozzi

Classifications

• G06N3/047

  Probabilistic or stochastic networks · CPC title

• G06N3/045 Primary

  Combinations of networks · CPC title

• G06N3/088 Primary

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

• G10L15/16

  using artificial neural networks · CPC title

• G06N3/08 Primary

  Learning methods · CPC title