Model generation apparatus, regression apparatus, model generation method, and computer-readable storage medium storing a model generation program
A model generation apparatus trains, through machine learning, a neural network module that includes an extraction operation to extract an element satisfying a predetermined condition from a set of targets. In the machine learning, the model generation apparatus performs the extraction operation in a phase of forward propagation with the neural network module, and replaces, in a phase of backpropagation, the extraction operation with a differentiable alternative operation and differentiates the alternative operation to compute an approximate gradient corresponding to differentiation for the extraction operation.
1. A model generation apparatus comprising a processor configured with a program to perform operations comprising:
operation as a data obtainer configured to obtain a plurality of training data pairs each comprising a combination of one or more training images and true information indicating ground truth for a real number to be computed using regression from the one or more training images; and
operation as a training unit configured to train a neural network module through machine learning using the plurality of training data pairs, wherein
the neural network module comprises an extraction operation to extract an element satisfying a predetermined condition from a set of targets,
the machine learning comprises training the neural network module on each of the plurality of training data pairs to compute, from the one or more training images using regression, a value fitting the ground truth indicated by the true information, and
the processor is configured with the program to perform operations such that operation as the training unit comprises performing, in the machine learning, the extraction operation in a phase of forward propagation with the neural network module and replacing, in a phase of backpropagation, the extraction operation with a differentiable alternative operation, and differentiating the alternative operation to compute an approximate gradient corresponding to differentiation for the extraction operation.
2. The model generation apparatus according to claim 1 , wherein
the neural network module further comprises an operation to compute a likelihood of each of a plurality of candidate values being a real number to be computed using regression, and
the extraction operation comprises extracting an element of a maximum from the computed likelihoods.
3. The model generation apparatus according to claim 1 , wherein
the one or more training images comprises a plurality of images captured at different positions, and
the real number to be computed using regression is a disparity between the plurality of images.
4. The model generation apparatus according to claim 1 , wherein
the one or more training images comprises a plurality of images captured continuously, and
the real number to be computed using regression is an estimate for a motion of an object in the plurality of images.
5. The model generation apparatus according to claim 1 , wherein
the one or more training images comprises images of an object, and
the real number to be computed using regression is an estimate for a detected position of the object.
6. A regression apparatus comprising a processor configured with a program to perform operations comprising:
operation as an image obtainer configured to obtain one or more target images;
operation as a regression unit configured to compute a real number using regression from the obtained one or more target images with a neural network module having been trained thereby; and
operation as an output unit configured to output information about a regression result, wherein
the neural network module is trained and generated through machine learning using a plurality of training data pairs each comprising a combination of one or more training images and true information indicating ground truth for a real number to be computed using regression from the one or more training images,
the neural network module comprises an extraction operation to extract an element satisfying a predetermined condition from a set of targets,
the machine learning comprises training the neural network module on each of the plurality of training data pairs to compute, from the one or more training images using regression, a value fitting the ground truth indicated by the true information, and
the processor is configured with the program to perform operations such that operation as the regression unit comprises performing, in the machine learning, the extraction operation in a phase of forward propagation with the neural network module and replacing, in a phase of backpropagation, the extraction operation with a differentiable alternative operation, and differentiating the alternative operation to compute an approximate gradient corresponding to differentiation for the extraction operation.
7. A model generation method implementable with a computer, the method comprising:
obtaining a plurality of training data pairs each comprising a combination of one or more training images and true information indicating ground truth for a real number to be computed using regression from the one or more training images; and
training a neural network module through machine learning using the plurality of training data pairs, wherein
the neural network module comprises an extraction operation to extract an element satisfying a predetermined condition from a set of targets,
the machine including comprises training the neural network module on each of the plurality of training data pairs to compute, from the one or more training images using regression, a value fitting the ground truth indicated by the true information, and
training the neural network module comprises performing, in the machine learning, the extraction operation in a phase of forward propagation with the neural network module, and replacing, in a phase of backpropagation, the extraction operation with a differentiable alternative operation, and differentiating the alternative operation to compute an approximate gradient corresponding to differentiation for the extraction operation.
8. The model generation method according to claim 7 , wherein
the neural network module further comprises an operation to compute a likelihood of each of a plurality of candidate values being a real number to be computed using regression, and
the extraction operation comprises extracting an element of a maximum from the computed likelihoods.
9. A non-transitory computer-readable storage medium storing a model generation program, which when read and executed, causes a computer to perform operations comprising:
obtaining a plurality of training data pairs each including a combination of one or more training images and true information indicating ground truth for a real number to be computed using regression from the one or more training images; and
training a neural network module through machine learning using the plurality of training data pairs, wherein
the neural network module comprises an extraction operation to extract an element satisfying a predetermined condition from a set of targets,
the machine learning comprises training the neural network module on each of the plurality of training data pairs to compute, from the one or more training images using regression, a value fitting the ground truth indicated by the true information,
the training the neural network module comprises performing, in the machine learning, the extraction operation in a phase of forward propagation with the neural network module, and replacing, in a phase of backpropagation, the extraction operation with a differentiable alternative operation, and differentiating the alternative operation to compute an approximate gradient corresponding to differentiation for the extraction operation.
10. The model generation program according to claim 9 , wherein
the neural network module further comprises an operation to compute a likelihood of each of a plurality of candidate values being a real number to be computed using regression, and
the extraction operation comprises extracting an element of a maximum from the computed likelihoods.
11. The model generation apparatus according to claim 2 , wherein
the one or more training images comprise a plurality of images captured at different positions, and
the real number to be computed using regression is a disparity between the plurality of images.
12. The model generation apparatus according to claim 2 , wherein
the one or more training images comprise a plurality of images captured continuously, and
the real number to be computed using regression is an estimate for a motion of an object in the plurality of images.
13. The model generation apparatus according to claim 2 , wherein
the one or more training images comprise images of an object, and
the real number to be computed using regression is an estimate for a detected position of the object.