System and method for semi-supervised conditional generative modeling using adversarial networks
One embodiment facilitates generating synthetic data objects using a semi-supervised GAN. During operation, a generator module synthesizes a data object derived from a noise vector and an attribute label. The system passes, to an unsupervised discriminator module, the data object and a set of training objects which are obtained from a training data set. The unsupervised discriminator module calculates: a value indicating a probability that the data object is real; and a latent feature representation of the data object. The system passes the latent feature representation and the attribute label to a supervised discriminator module. The supervised discriminator module calculates a value indicating a probability that the attribute label given the data object is real. The system performs the aforementioned steps iteratively until the generator module produces data objects with a given attribute label which the unsupervised and supervised discriminator modules can no longer identify as fake.
1. A computer-implemented method for generating synthetic data objects using a semi-supervised generative adversarial network, the method comprising:
synthesizing, by a generator module, a data object x G derived from a noise vector z and an attribute label y,
wherein the semi-supervised generative adversarial network comprises the generator module, an unsupervised discriminator module, and a supervised discriminator module;
passing, to the unsupervised discriminator module, the data object x G and a set of training objects x T and x U which are obtained from a training data set,
wherein the training data set includes the x U objects which do not have a corresponding attribute label and further includes the x T objects that do have corresponding attribute labels y T ;
calculating, by the unsupervised discriminator module, a value indicating a probability that the data object x G is real;
producing, by the unsupervised discriminator module, a first latent feature representation h(x G ) of the data object x G and a second latent feature representation h(x T ) of the data object x T ;
passing the first and second latent feature representations h(x G ) and h(x T ) to the supervised discriminator module;
passing the attribute label y and an attribute label y T corresponding to the data object x T to the supervised discriminator module;
receiving, by the supervised discriminator module as input, a first pair comprising the first latent feature representation h(x G ) produced by the unsupervised discriminator module and the attribute label y and a second pair comprising the second latent feature representation h(x T ) produced by the unsupervised discriminator module and the corresponding y T attribute;
calculating, by the supervised discriminator module based on at least the first pair and the second pair, a value indicating a probability that the attribute label y given the data object x G is real; and
performing the aforementioned steps iteratively until the generator module produces data objects with a given attribute label which the unsupervised and supervised discriminator modules can no longer identify as fake.
2. The method of claim 1 , wherein calculating the value indicating the probability that the data object x G is real further comprises determining that the data object x G is obtained from the training data set.
3. The method of claim 1 , wherein the training data set includes a first number of data objects which do not have a corresponding attribute label and a second number of data objects which do have a corresponding attribute label, and wherein the first number is greater by a predetermined ratio than the second number.
4. The method of claim 1 , wherein the generator module, the unsupervised discriminator module, and the supervised discriminator module are deep neural networks.
5. The method of claim 1 , wherein the generator module, the unsupervised discriminator module, and the supervised discriminator module comprise a model based on the semi-supervised generative adversarial network which:
learns a first probability that data objects are real based on data objects which have a corresponding attribute label and data objects which do not have a corresponding attribute label; and
learns a second probability that pairs comprised of a data object and a corresponding attribute label are real based on data objects which only have a corresponding attribute label; and
wherein the model subsequently uses a partially labeled given data set to determine a dependency between a given data object and a given attribute label of the given data set; and
subsequently generates a specific data object given a specific attribute label that satisfies the dependency between the given data object and the given attribute label.
6. The method of claim 1 , wherein a data object and a corresponding attribute label are one or more of:
an image and an attribute for the image;
an audio file and an attribute for the audio file; and
a first set of data and a tag for the first set of data.
7. The method of claim 1 , wherein a data object is an image of a face, and a corresponding attribute label for the data object pertains to a presence or an absence of one or more of sunglasses, wrinkles, and facial cosmetics.
8. The method of claim 1 , wherein a data object is an image of an alphanumeric character, and a corresponding attribute label for the data object pertains uniquely to the alphanumeric character.
9. A computer system for generating synthetic data objects using a semi-supervised generative adversarial network, the system comprising:
a processor; and
a storage device storing instructions that when executed by the processor cause the processor to perform a method, the method comprising:
synthesizing, by a generator module, a data object x G derived from a noise vector z and an attribute label y,
wherein the semi-supervised generative adversarial network comprises the generator module, an unsupervised discriminator module, and a supervised discriminator module;
passing, to the unsupervised discriminator module, the data object x G and a set of training objects x T and x U which are obtained from a training data set,
wherein the training data set includes the x U objects which do not have a corresponding attribute label and further includes the x T objects that do have corresponding attribute labels y T ;
calculating, by the unsupervised discriminator module, a value indicating a probability that the data object x G is real;
producing, by the unsupervised discriminator module, a first latent feature representation h(x G ) of the data object x G and a second latent feature representation h(x T ) of the data object x T ;
passing the first and second latent feature representations h(x G ) and h(x T ) to the supervised discriminator module;
passing the attribute label y and an attribute label y T corresponding to the data object x T to the supervised discriminator module;
receiving, by the supervised discriminator module as input, a first pair comprising the first latent feature representation h(x G ) produced by the unsupervised discriminator module and the attribute label y and a second pair comprising the second latent feature representation h(x T ) produced by the unsupervised discriminator module and the corresponding y T attribute;
calculating, by the supervised discriminator module based on at least the first pair and the second pair, a value indicating a probability that the attribute label y given the data object x G is real; and
performing the aforementioned steps iteratively until the generator module produces data objects with a given attribute label which the unsupervised and the supervised discriminator modules can no longer identify as fake.
10. The computer system of claim 9 , wherein calculating the value indicating the probability that the data object x G is real further comprises determining that the data object x G is obtained from the training data set.
11. The computer system of claim 9 , wherein the training data set includes a first number of data objects which do not have a corresponding attribute label and a second number of data objects which do have a corresponding attribute label, and wherein the first number is greater by a predetermined ratio than the second number.
12. The computer system of claim 9 , wherein the generator module, the unsupervised discriminator module, and the supervised discriminator module are deep neural networks.
13. The computer system of claim 9 , wherein the generator module, the unsupervised discriminator module, and the supervised discriminator module comprise a model based on the semi-supervised generative adversarial network which:
learns a first probability that data objects are real based on data objects which have a corresponding attribute label and data objects which do not have a corresponding attribute label; and
learns a second probability that pairs comprised of a data object and a corresponding attribute label are real based on data objects which only have a corresponding attribute label; and
wherein the model subsequently uses a partially labeled given data set to determine a dependency between a given data object and a given attribute label of the given data set; and
subsequently generates a specific data object given a specific attribute label that satisfies the dependency between the given data object and the given attribute label.
14. The computer system of claim 9 , wherein a data object and a corresponding attribute label are one or more of:
an image and an attribute for the image;
an audio file and an attribute for the audio file; and
a first set of data and a tag for the first set of data.
15. The computer system of claim 9 , wherein a data object is an image of a face, and a corresponding attribute label for the data object pertains to a presence or an absence of one or more of sunglasses, wrinkles, and facial cosmetics.
16. The computer system of claim 9 , wherein a data object is an image of an alphanumeric character, and a corresponding attribute label for the data object pertains uniquely to the alphanumeric character.