Information processing apparatus and information processing method
An information processing method includes: inputting sample image into a machine learning architecture to obtain a first feature, and causing a first classifier to calculate a first classification loss; calculating a second feature based on the first feature and a predetermined first mask, and inputting the second feature into the first classifier to calculate an entropy loss; calculating a second mask based on the first mask and the entropy loss to maximize the entropy loss; obtaining an adversarial feature based on the first feature and the second mask, where the adversarial feature is complementary to the second feature; causing, by training the first classifier and the second classifier in association with each other, the second classifier to calculate a second classification loss based on the adversarial feature; and adjusting parameters of the machine learning architecture, the first classifier and the second classifier, to obtain a trained machine learning architecture.
1. An information processing apparatus, comprising:
a first classification loss calculating unit configured to input each sample image in a training set into a machine learning architecture to obtain a first feature of the sample image, and to cause a first classifier to calculate a first classification loss of the sample image based on the first feature;
an entropy loss calculating unit configured to calculate a second feature based on the first feature and a predetermined first mask, and to input the second feature into the first classifier to calculate an entropy loss of class probability distribution;
a mask calculating unit configured to calculate a second mask based on the first mask and the entropy loss to maximize the entropy loss;
an adversarial feature obtaining unit configured to obtain an adversarial feature based on the first feature and the second mask, wherein the adversarial feature is complementary to the second feature;
a second classification loss calculating unit configured to cause, by training the first classifier and the second classifier in association with each other, the second classifier to calculate a second classification loss of the sample image based on the adversarial feature; and
a training unit configured to adjust, by minimizing a sum of the first classification loss and the second classification loss, parameters of the machine learning architecture, the first classifier and the second classifier, to obtain a trained machine learning architecture.
2. The information processing apparatus according to claim 1 , wherein the mask calculating unit is configured to calculate the second mask based on a gradient of the entropy loss with respect to the first mask.
3. The information processing apparatus according to claim 1 , wherein the mask calculating unit is configured to calculate, with a Long Short-Term Memory Model (LSTM), the second mask based on the entropy loss and the first mask.
4. The information processing apparatus according to claim 1 , wherein the training the first classifier and the second classifier in association with each other comprises: causing parameters of the second classifier to comprise parameters of the first classifier, and causing the second classifier to calculate the second classification loss of the sample image based only on the adversarial feature.
5. The information processing apparatus according to claim 1 , wherein the training the first classifier and the second classifier in association with each other comprises: causing parameters of the second classifier not to comprise parameters of the first classifier, and causing the second classifier to calculate the second classification loss of the sample image based on both the second feature and the adversarial feature.
6. The information processing apparatus according to claim 1 , wherein the adversarial feature obtaining unit is configured to average features obtained after multiplying the first feature with the second mask, to obtain the adversarial feature.
7. The information processing apparatus according to claim 1 , wherein the entropy loss calculating unit is configured to average features obtained after multiplying the first feature with the first mask, to obtain the second feature.
8. The information processing apparatus according to claim 1 , the first feature is a three-dimensional feature.
9. The information processing apparatus according to claim 1 , wherein the machine learning architecture is a neural network architecture.
10. An apparatus for performing classification with the trained machine learning architecture obtained by the information processing apparatus according to claim 1 , comprising:
a feature extracting unit configured to input an image to be classified into the trained machine learning architecture, to extract a feature of the image to be classified; and
a classifying unit configured to calculate feature similarities between the feature of the image to be classified and features of images of known categories, and to classify the image to be classified into a category to which an image of a known category having a largest feature similarity with the image to be classified belongs.
11. The apparatus according to claim 10 , wherein the classifying unit is configured to average the features of the image to be classified, and calculate the feature similarity based on the averaged features.
12. An information processing method, comprising:
a first classification loss calculating step of inputting ach sample image in a training set into a machine learning architecture to obtain a first feature of the sample image, and causing a first classifier to calculate a first classification loss of the sample image based on the first feature;
an entropy loss calculating step of calculating a second feature based on the first feature and a predetermined first mask, and inputting the second feature into the first classifier to calculate an entropy loss of class probability distribution;
a mask calculating step of calculating a second mask based on the first mask and the entropy loss to maximize the entropy loss;
an adversarial feature obtaining step of obtaining an adversarial feature based on the first feature and the second mask, wherein the adversarial feature is complementary to the second feature;
a second classification loss calculating step of causing, by training the first classifier and the second classifier in association with each other, the second classifier to calculate a second classification loss of the sample image based on the adversarial feature; and
to a training step of adjusting, by minimizing a sum of the first classification loss and the second classification loss, parameters of the machine learning architecture, the first classifier and the second classifier, to obtain a trained machine learning architecture.
13. The information processing method according to claim 12 , wherein in the mask calculating step, the second mask is calculated based on a gradient of the entropy loss with respect to the first mask.
14. The information processing method according to claim 12 , wherein in the mask calculating steps, the second mask is calculated based on the entropy loss and the first mask with a Long Short-Term Memory Model (LSTM).
15. The information processing method according to claim 12 , wherein the training the first classifier and the second classifier in association with each other comprises: causing parameters of the second classifier to comprise parameters of the first classifier, and causing the second classifier to calculate the second classification loss of the sample image based only on the adversarial feature.
16. The information processing method according to claim 12 , wherein the training the first classifier and the second classifier in association with each other comprises: causing parameters of the second classifier not to comprise parameters of the first classifier, and causing the second classifier to calculate the second classification loss of the sample image based on both the second feature and the adversarial feature.
17. The information processing method according to claim 12 , wherein in the adversarial feature obtaining step, features obtained after multiplying the first feature with the second mask are averaged, to obtain the adversarial feature.
18. The information processing method according to claim 12 , wherein in the entropy loss calculating step, features obtained after multiplying the first feature with the first mask are averaged, to obtain the second feature.
19. The information processing method according to claim 12 , wherein the first feature is a three-dimensional feature.
20. The information processing method according to claim 12 , wherein the machine learning architecture is a neural network architecture.