Monte-Carlo adversarial autoencoder for multi-source domain adaptation
Embodiments may include novel techniques for training and using an adversarial autoencoder for multi-source domain functions. For example, a method may comprise training an adversarial encoder comprising an encoder and a decoder by simultaneously training the encoder and the decoder, using data comprising a plurality of datasets, the data having labels based on an origin class and a dataset number, training the encoder to act as a generator to generate codewords based on the data for a generative adversarial network including the generator and a discriminator by training the generator to cause the discriminator to predict random labels for a plurality of data samples of each class and training the generator using the predicted random labels to generate codewords that relate to the origin class, and classifying new data samples using the trained adversarial encoder and generator, and the discriminator.
1 . A method, implemented in a computer system comprising a processor, memory accessible by the processor, and computer program instructions stored in the memory and executable by the processor, the method comprising:
training an adversarial encoder comprising an encoder and a decoder by simultaneously training the encoder and the decoder, using data comprising a plurality of datasets, the data having labels based on an origin class and a dataset number;
training the encoder to act as a generator to generate codewords based on the data for a generative adversarial network including the generator and a discriminator by training the generator to cause the discriminator to predict random labels for a plurality of data samples of each class and training the generator using the predicted random labels to generate codewords that relate to the origin class;
wherein the random labels are constantly changed utilizing a Monte-Carlo method and mapped based on the codewords; and
classifying new data samples using the trained adversarial encoder and generator, and the discriminator by utilizing a Generative Adversarial Networks (GAN) for domain adaptation and discrimination between a plurality of source labels to enforce a plurality of constraints on the codewords;
wherein each codeword represents a behavior pattern that is class aware and distinguishes between the plurality of source labels.
2 . The method of claim 1 , wherein the generated codewords map items with a correct label from all sources utilizing the adversarial encoder comprising Generative Adversarial Networks (GANs).
3 . The method of claim 1 , wherein the discriminator is a label/source discriminator that distinguishes between exact origin classes.
4 . The method of claim 3 , wherein the label/source discriminator is used with the decoder to group similar data samples together and to make each label indistinguishable between the origin classes.
5 . The method of claim 1 , wherein new data samples are classified by:
encoding the new data sample using the adversarial encoder to form a codeword;
training a classifier using a set of all codewords and a set of all labels; and
classifying the new data sample using the trained classifier and the formed codeword.
6 . A system comprising a processor, memory accessible by the processor, and computer program instructions stored in the memory and executable by the processor to perform:
training an adversarial encoder comprising an encoder and a decoder by simultaneously training the encoder and the decoder, using data comprising a plurality of datasets, the data having labels based on an origin class and a dataset number;
training the encoder to act as a generator to generate codewords based on the data for a generative adversarial network including the generator and a discriminator by training the generator to cause the discriminator to predict random labels for a plurality of data samples of each class and training the generator using the predicted random labels to generate codewords that relate to the origin class;
wherein the random labels are constantly changed utilizing a Monte-Carlo method and mapped based on the codewords; and
classifying new data samples using the trained adversarial encoder and generator, and the discriminator by utilizing a Generative Adversarial Networks (GAN) for domain adaptation and discrimination between a plurality of source labels to enforce a plurality of constraints on the codewords;
wherein each codeword represents a behavior pattern that is class aware and distinguishes between the plurality of source labels.
7 . The system of claim 6 , wherein the generated codewords map items with a correct label from all sources utilizing the adversarial encoder comprising Generative Adversarial Networks (GANs).
8 . The system of claim 6 , wherein the discriminator is a label/source discriminator that distinguishes between exact origin classes.
9 . The system of claim 8 , wherein the label/source discriminator is used with the decoder to group similar data samples together and to make each label indistinguishable between the origin classes.
10 . The system of claim 6 , wherein new data samples are classified by:
encoding the new data sample using the adversarial encoder to form a codeword;
training a classifier using a set of all codewords and a set of all labels; and
classifying the new data sample using the trained classifier and the formed codeword.
11 . A computer program product comprising a non-transitory computer readable storage having program instructions embodied therewith, the program instructions executable by a computer, to cause the computer to perform a method comprising:
training an adversarial encoder comprising an encoder and a decoder by simultaneously training the encoder and the decoder, using data comprising a plurality of datasets, the data having labels based on an origin class and a dataset number;
training the encoder to act as a generator to generate codewords based on the data for a generative adversarial network including the generator and a discriminator by training the generator to cause the discriminator to predict random labels for a plurality of data samples of each class and training the generator using the predicted random labels to generate codewords that relate to the origin class;
wherein the random labels are constantly changed utilizing a Monte-Carlo method and mapped based on the codewords; and
classifying new data samples using the trained adversarial encoder and generator, and the discriminator by utilizing a Generative Adversarial Networks (GAN) for domain adaptation and discrimination between a plurality of source labels to enforce a plurality of constraints the codewords;
wherein each codeword represents a behavior pattern that is class aware and distinguishes between the plurality of source labels.
12 . The computer program product of claim 11 , wherein the generated codewords map items with a correct label from all sources utilizing the adversarial encoder comprising Generative Adversarial Networks (GANs).
13 . The computer program product of claim 11 , wherein the discriminator is a label/source discriminator that distinguishes between exact origin classes.
14 . The computer program product of claim 13 , wherein the label/source discriminator is used with the decoder to group similar data samples together and to make each label indistinguishable between the origin classes.
15 . The computer program product of claim 11 , wherein new data samples are classified by:
encoding the new data sample using the adversarial encoder to form a codeword;
training a classifier using a set of all codewords and a set of all labels; and
classifying the new data sample using the trained classifier and the formed codeword.