Information processing apparatus, information processing method, and non-transitory computer-readable storage medium
An information processing apparatus comprises a generating unit configured to generate supervised data relating to a topological invariant based on a topological property relating to supervised data corresponding to input data, and a training unit configured to perform, based on a geometric property relating to an output from a classifier to which the input data is input and the supervised data generated by the generating unit, training of the classifier.
1. An information processing apparatus comprising:
a memory storing instructions; and
a processor that executes the instructions to:
obtain an output image output from a classifier that receives input data;
calculate a first error between the output image and a supervised image corresponding to the input data;
set a region of interest in the supervised image for calculating a topological invariant, which is a property that is kept unchanged even if an original shape is continuously transformed;
calculate a first topological invariant of the region of interest in the supervised image;
calculate a second topological invariant of a correspondence region in the output image corresponding to the region of interest;
calculate a second error between the first topological invariant and the second topological invariant;
calculate a total error from the first error and the second error; and
train the classifier to minimize the total error.
2. The information processing apparatus according to claim 1 , wherein the processor, in setting the region of interest, sets a size and a shape of the region of interest based on the first topological invariant.
3. The information processing apparatus according to claim 1 , wherein:
the processor executes instructions to display the supervised image on a display, and
the processor, in setting the region of interest, sets a size and a shape of the region of interest based on an input by a user in relation to the displayed supervised image.
4. The information processing apparatus according to claim 1 , the processor executes the instructions to:
obtain training data to be used in retraining of the classifier from a set of training data including the input data and the supervised image; and
retrain the classifier using the obtained training data.
5. The information processing apparatus according to claim 4 , wherein the processor obtains the training data to be used in the retraining of the classifier from the set based on a result of the training of the classifier.
6. The information processing apparatus according to claim 5 , wherein the processor, in obtaining the training data, selects from the set, as the training data to be used in the retraining of the classifier, a predetermined number of training data in order of increasing error in the result of the training of the classifier.
7. The information processing apparatus according to claim 5 , wherein the processor, in obtaining the training data, displays the set in a list and obtains selected training data selected by a user from the set displayed in the list as the training data to be used in the retraining of the classifier.
8. The information processing apparatus according to claim 1 , wherein the processor calculates the second error based on a cross entropy where the first and second topological invariants each are a label value.
9. The information processing apparatus according to claim 1 , wherein the classifier is a hierarchical neural network.
10. The information processing apparatus according to claim 1 , wherein:
the first topological invariant is a number of holes that exists in a contour of the region of interest in the supervised image, and
the second topological invariant is a number of holes that exists in a contour of the correspondence region in the output image.
11. An information processing method, the method comprising:
obtaining an output image output from a classifier that receives input data;
calculating a first error between the output image and a supervised image corresponding to the input data;
setting a region of interest in the supervised image for calculating a topological invariant, which is a property that is kept unchanged even if an original shape is continuously transformed;
calculating a first topological invariant of the region of interest in the supervised imago;
calculating a second topological invariant of a correspondence region in the output image corresponding to the region of interest;
calculating a second error between the first topological invariant and the second topological invariant;
calculating a total error from the first error and the second error; and
training the classifier to minimize the total error.
12. A non-transitory computer-readable storage medium storing a computer program executable by a computer to execute a method comprising:
obtaining an output image output from a classifier that receives input data;
calculating a first error between the output image and a supervised image corresponding to the input data;
setting a region of interest in the supervised image for calculating a topological invariant, which is a property that is kept unchanged even if an original shape is continuously transformed;
calculating a first topological invariant of the region of interest in the supervised image;
calculating a second topological invariant of a correspondence region in the output image corresponding to the region of interest;
calculating a second error between the first topological invariant and the second topological invariant;
calculating a total error from the first error and the second error; and
training the classifier to minimize the total error.