Patent
US 11,263,487Multi-Task GAN, and Image Translator and Image Classifier Trained Thereby
Qun LI
US
Copyright © 2026 Matter42. All rights reserved.
Patent
Atlas literature
Patent
US 11,263,487Patent drawings and their descriptions. Click a drawing to enlarge it.
Claims define the patent's legal scope. Independent claims stand alone; dependent claims (nested) narrow them. Click a claim to expand its dependents.
One or more computing devices for performing machine training, comprising: a conditional generative adversarial network (GAN) including a generator neural network and a discriminator neural network, the generator neural network being configured to: receive generator input information that includes a conditional input image and plural conditional input values; and transform the generator input information into a generator output image, the plural conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, the discriminator neural network being configured to: receive discriminator input information that includes the conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural attribute values based on the discriminator input information, each attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural attribute values being associated with the same plural attribute types as the plural conditional input values; and produce an output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or the real image; and a parameter-updating system for iteratively adjusting parameter values of the generator neural network and the discriminator neural network, upon completion of training, the parameter-updating system providing a first trained model based on trained parameter values associated with the generator neural network, and a second trained model based on trained parameter values associated with the discriminator neural network for use by an inference- stage image classifier without the first trained model, the G AN and the parameter-updating system being implemented by hardware logic circuitry provided by said one or more computing devices. Currently amended
The one or more computing devices of claim 1, wherein the conditional input image is a transformed version of the real image. Original
The one or more computing devices of claim 1, wherein the discriminator neural network includes a convolutional neural network for mapping the discriminator input information into feature information, and plural individual classifier neural networks for respectively producing the plural attribute values and the output result that conveys whether the discriminator input information includes the generator output image or the real image. Original
The one or more computing devices of claim 1, wherein the inference- stage image classifier is configured to receive classifier input information that includes a classifier input image and a classifier conditional input image, the classifier conditional input image being generated by transforming the classifier input image without use of the first trained model. New
The one or more computing devices of claim 1, wherein the training involves increasing a number of training examples fed to the discriminator neural network by providing different sets of attribute values to the generator neural network for the same conditional input image. New
The one or more computing devices of claim 1, wherein the discriminator neural network is configured to reduce a number of channels in the discriminator input information before further processing. New
A method for training a conditional generative adversarial network (G AN), including: using a generator neural network of the G AN to: receive generator input information that includes a conditional input image and plural conditional input values; and transform the generator input information into a generator output image, the plural conditional input conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, using a discriminator neural network of the G AN to: receive discriminator input information that includes the conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural attribute values based on the discriminator input information, each attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural attribute values being associated with the same plural attribute types as the plural conditional input values; and produce an output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or the real image; and iteratively adjusting parameter values of the generator neural network and the discriminator neural network, upon completion of training, the method providing a first trained model based on trained parameter values associated with the generator neural network, and a second trained model based on trained parameter values associated with the discriminator neural network for use by an inference-stage image classifier without the use of the first trained model. Currently amended
The method of claim 4, wherein the conditional input image is a transformed version of the real image. Original
The method of claim 4, wherein the discriminator neural network includes a convolutional neural network for mapping the discriminator input information into feature information, and plural individual classifier neural networks for respectively producing the plural attribute values and the output result that conveys whether the discriminator input information includes the generator output image or the real image. Original
The method of claim 4, further comprising using the inference-stage image classifier by: receiving classifier input information that includes a classifier input image and a classifier conditional input image, the classifier conditional input image being generated by transforming the classifier input image without use of the first trained model; producing plural classifier attribute values based on the classifier input information, each classifier attribute value being associated with a characteristic of an object depicted by the classifier input image; and producing a classifier output result based on the classifier input information that indicates whether the translator input image is synthetic or real. Currently amended
An image translator produced by the method of claim 4. Original
An image classifier produced by the method of claim 4. Original
The method of claim 4, wherein the training involves increasing a number of training examples fed to the discriminator neural network by providing different sets of attribute values to the generator neural network for the same conditional input image. New
. Canceled
. Canceled
. Canceled
The method of clam 11, wherein the supplemental content item is presented in the electronic document in which the classifier input image appears. Original
An image classification system implemented by one or more computing devices, for operation in an inference stage, comprising: hardware logic circuitry configured to: receive a classifier input image to be classified; transform the classifier input image into a classifier conditional input image, the classifier input image and the classifier conditional input image corresponding to classifier input information; use an image classifier neural network provided by the hardware logic circuitry to produce plural classifier attribute values based on the classifier input information, each classifier attribute value being associated with a characteristic of an object depicted by the classifier input image; and use the image classifier neural network to produce a classifier output result based on the classifier input information that indicates whether the classifier input image is synthetic or real, the image classifier neural network having a model that is trained, in a prior training process, by iteratively adjusting parameter values of a discriminator neural network in a generative adversarial network (GAN), wherein the GAN includes a generator neural network in addition to the discriminator neural network, the generator neural network being configured to: receive generator input information that includes a generator conditional input image and plural generator conditional input values; and transform the generator input information into a generator output image, the plural generator conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, the discriminator neural network being configured to: receive discriminator input information that includes the generator conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural discriminator attribute values based on the discriminator input information, each discriminator attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural discriminator attribute values being associated with the same plural attribute types as the plural generator conditional input values; and produce a discriminator output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or a real image. Currently amended
The image classification system of claim 15, wherein the classifier neural network includes a convolutional neural network for mapping the classifier input information into feature information, and plural individual classifier neural networks for respectively producing the plural classifier attribute values and the classifier output result that conveys whether the classifier input information is synthetic or real. Original
The image classification system of claim 15, wherein the plural attribute types include any two or more attribute types selected from: a category attribute type; a color attribute type; a department attribute type; a material attribute type; and a pattern attribute type. Currently amended
The image classification system of claim 15, wherein the classifier input image originates from an electronic document with which a user is interacting via a user computing device. Original
. Canceled
The image classification system of claim [[16]] 15, wherein each generator conditional input image fed to the generator neural network and the discriminator neural network is produced by a same image transformation that is used to produce the classifier conditional input image. Currently amended
Layer stacks claimed or described, ordered top of device to substrate.
conditional generative adversarial network (GAN) training system
No layer stack recorded.
image classification system (inference stage)
No layer stack recorded.
Patent
Atlas literature
Patent
US 11,263,487Patent drawings and their descriptions. Click a drawing to enlarge it.
Claims define the patent's legal scope. Independent claims stand alone; dependent claims (nested) narrow them. Click a claim to expand its dependents.
One or more computing devices for performing machine training, comprising: a conditional generative adversarial network (GAN) including a generator neural network and a discriminator neural network, the generator neural network being configured to: receive generator input information that includes a conditional input image and plural conditional input values; and transform the generator input information into a generator output image, the plural conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, the discriminator neural network being configured to: receive discriminator input information that includes the conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural attribute values based on the discriminator input information, each attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural attribute values being associated with the same plural attribute types as the plural conditional input values; and produce an output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or the real image; and a parameter-updating system for iteratively adjusting parameter values of the generator neural network and the discriminator neural network, upon completion of training, the parameter-updating system providing a first trained model based on trained parameter values associated with the generator neural network, and a second trained model based on trained parameter values associated with the discriminator neural network for use by an inference- stage image classifier without the first trained model, the G AN and the parameter-updating system being implemented by hardware logic circuitry provided by said one or more computing devices. Currently amended
The one or more computing devices of claim 1, wherein the conditional input image is a transformed version of the real image. Original
The one or more computing devices of claim 1, wherein the discriminator neural network includes a convolutional neural network for mapping the discriminator input information into feature information, and plural individual classifier neural networks for respectively producing the plural attribute values and the output result that conveys whether the discriminator input information includes the generator output image or the real image. Original
The one or more computing devices of claim 1, wherein the inference- stage image classifier is configured to receive classifier input information that includes a classifier input image and a classifier conditional input image, the classifier conditional input image being generated by transforming the classifier input image without use of the first trained model. New
The one or more computing devices of claim 1, wherein the training involves increasing a number of training examples fed to the discriminator neural network by providing different sets of attribute values to the generator neural network for the same conditional input image. New
The one or more computing devices of claim 1, wherein the discriminator neural network is configured to reduce a number of channels in the discriminator input information before further processing. New
A method for training a conditional generative adversarial network (G AN), including: using a generator neural network of the G AN to: receive generator input information that includes a conditional input image and plural conditional input values; and transform the generator input information into a generator output image, the plural conditional input conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, using a discriminator neural network of the G AN to: receive discriminator input information that includes the conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural attribute values based on the discriminator input information, each attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural attribute values being associated with the same plural attribute types as the plural conditional input values; and produce an output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or the real image; and iteratively adjusting parameter values of the generator neural network and the discriminator neural network, upon completion of training, the method providing a first trained model based on trained parameter values associated with the generator neural network, and a second trained model based on trained parameter values associated with the discriminator neural network for use by an inference-stage image classifier without the use of the first trained model. Currently amended
The method of claim 4, wherein the conditional input image is a transformed version of the real image. Original
The method of claim 4, wherein the discriminator neural network includes a convolutional neural network for mapping the discriminator input information into feature information, and plural individual classifier neural networks for respectively producing the plural attribute values and the output result that conveys whether the discriminator input information includes the generator output image or the real image. Original
The method of claim 4, further comprising using the inference-stage image classifier by: receiving classifier input information that includes a classifier input image and a classifier conditional input image, the classifier conditional input image being generated by transforming the classifier input image without use of the first trained model; producing plural classifier attribute values based on the classifier input information, each classifier attribute value being associated with a characteristic of an object depicted by the classifier input image; and producing a classifier output result based on the classifier input information that indicates whether the translator input image is synthetic or real. Currently amended
An image translator produced by the method of claim 4. Original
An image classifier produced by the method of claim 4. Original
The method of claim 4, wherein the training involves increasing a number of training examples fed to the discriminator neural network by providing different sets of attribute values to the generator neural network for the same conditional input image. New
. Canceled
. Canceled
. Canceled
The method of clam 11, wherein the supplemental content item is presented in the electronic document in which the classifier input image appears. Original
An image classification system implemented by one or more computing devices, for operation in an inference stage, comprising: hardware logic circuitry configured to: receive a classifier input image to be classified; transform the classifier input image into a classifier conditional input image, the classifier input image and the classifier conditional input image corresponding to classifier input information; use an image classifier neural network provided by the hardware logic circuitry to produce plural classifier attribute values based on the classifier input information, each classifier attribute value being associated with a characteristic of an object depicted by the classifier input image; and use the image classifier neural network to produce a classifier output result based on the classifier input information that indicates whether the classifier input image is synthetic or real, the image classifier neural network having a model that is trained, in a prior training process, by iteratively adjusting parameter values of a discriminator neural network in a generative adversarial network (GAN), wherein the GAN includes a generator neural network in addition to the discriminator neural network, the generator neural network being configured to: receive generator input information that includes a generator conditional input image and plural generator conditional input values; and transform the generator input information into a generator output image, the plural generator conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, the discriminator neural network being configured to: receive discriminator input information that includes the generator conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural discriminator attribute values based on the discriminator input information, each discriminator attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural discriminator attribute values being associated with the same plural attribute types as the plural generator conditional input values; and produce a discriminator output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or a real image. Currently amended
The image classification system of claim 15, wherein the classifier neural network includes a convolutional neural network for mapping the classifier input information into feature information, and plural individual classifier neural networks for respectively producing the plural classifier attribute values and the classifier output result that conveys whether the classifier input information is synthetic or real. Original
The image classification system of claim 15, wherein the plural attribute types include any two or more attribute types selected from: a category attribute type; a color attribute type; a department attribute type; a material attribute type; and a pattern attribute type. Currently amended
The image classification system of claim 15, wherein the classifier input image originates from an electronic document with which a user is interacting via a user computing device. Original
. Canceled
The image classification system of claim [[16]] 15, wherein each generator conditional input image fed to the generator neural network and the discriminator neural network is produced by a same image transformation that is used to produce the classifier conditional input image. Currently amended
Layer stacks claimed or described, ordered top of device to substrate.
conditional generative adversarial network (GAN) training system
No layer stack recorded.
image classification system (inference stage)
No layer stack recorded.
Patent
Atlas literature
Patent
US 11,263,487Patent drawings and their descriptions. Click a drawing to enlarge it.
Claims define the patent's legal scope. Independent claims stand alone; dependent claims (nested) narrow them. Click a claim to expand its dependents.
One or more computing devices for performing machine training, comprising: a conditional generative adversarial network (GAN) including a generator neural network and a discriminator neural network, the generator neural network being configured to: receive generator input information that includes a conditional input image and plural conditional input values; and transform the generator input information into a generator output image, the plural conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, the discriminator neural network being configured to: receive discriminator input information that includes the conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural attribute values based on the discriminator input information, each attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural attribute values being associated with the same plural attribute types as the plural conditional input values; and produce an output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or the real image; and a parameter-updating system for iteratively adjusting parameter values of the generator neural network and the discriminator neural network, upon completion of training, the parameter-updating system providing a first trained model based on trained parameter values associated with the generator neural network, and a second trained model based on trained parameter values associated with the discriminator neural network for use by an inference- stage image classifier without the first trained model, the G AN and the parameter-updating system being implemented by hardware logic circuitry provided by said one or more computing devices. Currently amended
The one or more computing devices of claim 1, wherein the conditional input image is a transformed version of the real image. Original
The one or more computing devices of claim 1, wherein the discriminator neural network includes a convolutional neural network for mapping the discriminator input information into feature information, and plural individual classifier neural networks for respectively producing the plural attribute values and the output result that conveys whether the discriminator input information includes the generator output image or the real image. Original
The one or more computing devices of claim 1, wherein the inference- stage image classifier is configured to receive classifier input information that includes a classifier input image and a classifier conditional input image, the classifier conditional input image being generated by transforming the classifier input image without use of the first trained model. New
The one or more computing devices of claim 1, wherein the training involves increasing a number of training examples fed to the discriminator neural network by providing different sets of attribute values to the generator neural network for the same conditional input image. New
The one or more computing devices of claim 1, wherein the discriminator neural network is configured to reduce a number of channels in the discriminator input information before further processing. New
A method for training a conditional generative adversarial network (G AN), including: using a generator neural network of the G AN to: receive generator input information that includes a conditional input image and plural conditional input values; and transform the generator input information into a generator output image, the plural conditional input conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, using a discriminator neural network of the G AN to: receive discriminator input information that includes the conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural attribute values based on the discriminator input information, each attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural attribute values being associated with the same plural attribute types as the plural conditional input values; and produce an output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or the real image; and iteratively adjusting parameter values of the generator neural network and the discriminator neural network, upon completion of training, the method providing a first trained model based on trained parameter values associated with the generator neural network, and a second trained model based on trained parameter values associated with the discriminator neural network for use by an inference-stage image classifier without the use of the first trained model. Currently amended
The method of claim 4, wherein the conditional input image is a transformed version of the real image. Original
The method of claim 4, wherein the discriminator neural network includes a convolutional neural network for mapping the discriminator input information into feature information, and plural individual classifier neural networks for respectively producing the plural attribute values and the output result that conveys whether the discriminator input information includes the generator output image or the real image. Original
The method of claim 4, further comprising using the inference-stage image classifier by: receiving classifier input information that includes a classifier input image and a classifier conditional input image, the classifier conditional input image being generated by transforming the classifier input image without use of the first trained model; producing plural classifier attribute values based on the classifier input information, each classifier attribute value being associated with a characteristic of an object depicted by the classifier input image; and producing a classifier output result based on the classifier input information that indicates whether the translator input image is synthetic or real. Currently amended
An image translator produced by the method of claim 4. Original
An image classifier produced by the method of claim 4. Original
The method of claim 4, wherein the training involves increasing a number of training examples fed to the discriminator neural network by providing different sets of attribute values to the generator neural network for the same conditional input image. New
. Canceled
. Canceled
. Canceled
The method of clam 11, wherein the supplemental content item is presented in the electronic document in which the classifier input image appears. Original
An image classification system implemented by one or more computing devices, for operation in an inference stage, comprising: hardware logic circuitry configured to: receive a classifier input image to be classified; transform the classifier input image into a classifier conditional input image, the classifier input image and the classifier conditional input image corresponding to classifier input information; use an image classifier neural network provided by the hardware logic circuitry to produce plural classifier attribute values based on the classifier input information, each classifier attribute value being associated with a characteristic of an object depicted by the classifier input image; and use the image classifier neural network to produce a classifier output result based on the classifier input information that indicates whether the classifier input image is synthetic or real, the image classifier neural network having a model that is trained, in a prior training process, by iteratively adjusting parameter values of a discriminator neural network in a generative adversarial network (GAN), wherein the GAN includes a generator neural network in addition to the discriminator neural network, the generator neural network being configured to: receive generator input information that includes a generator conditional input image and plural generator conditional input values; and transform the generator input information into a generator output image, the plural generator conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, the discriminator neural network being configured to: receive discriminator input information that includes the generator conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural discriminator attribute values based on the discriminator input information, each discriminator attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural discriminator attribute values being associated with the same plural attribute types as the plural generator conditional input values; and produce a discriminator output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or a real image. Currently amended
The image classification system of claim 15, wherein the classifier neural network includes a convolutional neural network for mapping the classifier input information into feature information, and plural individual classifier neural networks for respectively producing the plural classifier attribute values and the classifier output result that conveys whether the classifier input information is synthetic or real. Original
The image classification system of claim 15, wherein the plural attribute types include any two or more attribute types selected from: a category attribute type; a color attribute type; a department attribute type; a material attribute type; and a pattern attribute type. Currently amended
The image classification system of claim 15, wherein the classifier input image originates from an electronic document with which a user is interacting via a user computing device. Original
. Canceled
The image classification system of claim [[16]] 15, wherein each generator conditional input image fed to the generator neural network and the discriminator neural network is produced by a same image transformation that is used to produce the classifier conditional input image. Currently amended
Layer stacks claimed or described, ordered top of device to substrate.
conditional generative adversarial network (GAN) training system
No layer stack recorded.
image classification system (inference stage)
No layer stack recorded.
Patent
Atlas literature
Patent
US 11,263,487Patent drawings and their descriptions. Click a drawing to enlarge it.
Claims define the patent's legal scope. Independent claims stand alone; dependent claims (nested) narrow them. Click a claim to expand its dependents.
One or more computing devices for performing machine training, comprising: a conditional generative adversarial network (GAN) including a generator neural network and a discriminator neural network, the generator neural network being configured to: receive generator input information that includes a conditional input image and plural conditional input values; and transform the generator input information into a generator output image, the plural conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, the discriminator neural network being configured to: receive discriminator input information that includes the conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural attribute values based on the discriminator input information, each attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural attribute values being associated with the same plural attribute types as the plural conditional input values; and produce an output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or the real image; and a parameter-updating system for iteratively adjusting parameter values of the generator neural network and the discriminator neural network, upon completion of training, the parameter-updating system providing a first trained model based on trained parameter values associated with the generator neural network, and a second trained model based on trained parameter values associated with the discriminator neural network for use by an inference- stage image classifier without the first trained model, the G AN and the parameter-updating system being implemented by hardware logic circuitry provided by said one or more computing devices. Currently amended
The one or more computing devices of claim 1, wherein the conditional input image is a transformed version of the real image. Original
The one or more computing devices of claim 1, wherein the discriminator neural network includes a convolutional neural network for mapping the discriminator input information into feature information, and plural individual classifier neural networks for respectively producing the plural attribute values and the output result that conveys whether the discriminator input information includes the generator output image or the real image. Original
The one or more computing devices of claim 1, wherein the inference- stage image classifier is configured to receive classifier input information that includes a classifier input image and a classifier conditional input image, the classifier conditional input image being generated by transforming the classifier input image without use of the first trained model. New
The one or more computing devices of claim 1, wherein the training involves increasing a number of training examples fed to the discriminator neural network by providing different sets of attribute values to the generator neural network for the same conditional input image. New
The one or more computing devices of claim 1, wherein the discriminator neural network is configured to reduce a number of channels in the discriminator input information before further processing. New
A method for training a conditional generative adversarial network (G AN), including: using a generator neural network of the G AN to: receive generator input information that includes a conditional input image and plural conditional input values; and transform the generator input information into a generator output image, the plural conditional input conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, using a discriminator neural network of the G AN to: receive discriminator input information that includes the conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural attribute values based on the discriminator input information, each attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural attribute values being associated with the same plural attribute types as the plural conditional input values; and produce an output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or the real image; and iteratively adjusting parameter values of the generator neural network and the discriminator neural network, upon completion of training, the method providing a first trained model based on trained parameter values associated with the generator neural network, and a second trained model based on trained parameter values associated with the discriminator neural network for use by an inference-stage image classifier without the use of the first trained model. Currently amended
The method of claim 4, wherein the conditional input image is a transformed version of the real image. Original
The method of claim 4, wherein the discriminator neural network includes a convolutional neural network for mapping the discriminator input information into feature information, and plural individual classifier neural networks for respectively producing the plural attribute values and the output result that conveys whether the discriminator input information includes the generator output image or the real image. Original
The method of claim 4, further comprising using the inference-stage image classifier by: receiving classifier input information that includes a classifier input image and a classifier conditional input image, the classifier conditional input image being generated by transforming the classifier input image without use of the first trained model; producing plural classifier attribute values based on the classifier input information, each classifier attribute value being associated with a characteristic of an object depicted by the classifier input image; and producing a classifier output result based on the classifier input information that indicates whether the translator input image is synthetic or real. Currently amended
An image translator produced by the method of claim 4. Original
An image classifier produced by the method of claim 4. Original
The method of claim 4, wherein the training involves increasing a number of training examples fed to the discriminator neural network by providing different sets of attribute values to the generator neural network for the same conditional input image. New
. Canceled
. Canceled
. Canceled
The method of clam 11, wherein the supplemental content item is presented in the electronic document in which the classifier input image appears. Original
An image classification system implemented by one or more computing devices, for operation in an inference stage, comprising: hardware logic circuitry configured to: receive a classifier input image to be classified; transform the classifier input image into a classifier conditional input image, the classifier input image and the classifier conditional input image corresponding to classifier input information; use an image classifier neural network provided by the hardware logic circuitry to produce plural classifier attribute values based on the classifier input information, each classifier attribute value being associated with a characteristic of an object depicted by the classifier input image; and use the image classifier neural network to produce a classifier output result based on the classifier input information that indicates whether the classifier input image is synthetic or real, the image classifier neural network having a model that is trained, in a prior training process, by iteratively adjusting parameter values of a discriminator neural network in a generative adversarial network (GAN), wherein the GAN includes a generator neural network in addition to the discriminator neural network, the generator neural network being configured to: receive generator input information that includes a generator conditional input image and plural generator conditional input values; and transform the generator input information into a generator output image, the plural generator conditional input values describing plural respective characteristics of the generator output image associated with plural respective attribute types, the discriminator neural network being configured to: receive discriminator input information that includes the generator conditional input image and a discriminator input image, the discriminator input image corresponding to either the generator output image or a real image that is not generated by the generator neural network; produce plural discriminator attribute values based on the discriminator input information, each discriminator attribute value being associated with a characteristic of an object depicted by the discriminator input image, the plural discriminator attribute values being associated with the same plural attribute types as the plural generator conditional input values; and produce a discriminator output result based on the discriminator input information that indicates whether the discriminator input image is the generator output image or a real image. Currently amended
The image classification system of claim 15, wherein the classifier neural network includes a convolutional neural network for mapping the classifier input information into feature information, and plural individual classifier neural networks for respectively producing the plural classifier attribute values and the classifier output result that conveys whether the classifier input information is synthetic or real. Original
The image classification system of claim 15, wherein the plural attribute types include any two or more attribute types selected from: a category attribute type; a color attribute type; a department attribute type; a material attribute type; and a pattern attribute type. Currently amended
The image classification system of claim 15, wherein the classifier input image originates from an electronic document with which a user is interacting via a user computing device. Original
. Canceled
The image classification system of claim [[16]] 15, wherein each generator conditional input image fed to the generator neural network and the discriminator neural network is produced by a same image transformation that is used to produce the classifier conditional input image. Currently amended
Layer stacks claimed or described, ordered top of device to substrate.
conditional generative adversarial network (GAN) training system
No layer stack recorded.
image classification system (inference stage)
No layer stack recorded.