An information processing apparatus and information processing method

Abstract

The invention discloses an information processing apparatus and an information processing method. The information processing apparatus includes: a classification loss calculation unit that inputs eachsample image in a training set to at least two feature extraction sections to obtain at least two features of the sample image, and causes a classification section to calculate a classification lossof the sample image based on the at least two features; a sample pair extraction unit that extracts a plurality of sample pairs for calculating mutual information between each pair of features from each pair of features of at least one pair of features of the at least two features; an information loss calculation unit that inputs the plurality of sample pairs to the machine learning architecture corresponding to each pair of features and calculates information loss between each pair of features; and the joint training unit that is used for jointly adjusting the parameters of the at least two feature extraction components, the classification component and the machine learning architecture in a manner of minimizing the sum of the classification loss and the information loss of all the sampleimages in the training set so as to obtain the at least two trained feature extraction components and classification component.

Description

technical field [0001] The present disclosure relates to the field of image classification, and specifically relates to an information processing device and an information processing method capable of learning more generalized and robust features, and an information processing device and an information processing method capable of judging the difficulty of a training task . Background technique [0002] In the traditional image classification method, the gradient-based optimization method is used for training. However, the gradient-based optimization training method will reduce the amount of information of the extracted features during processing, so in some cases it will not learn All the discriminable features in the image, but only learn those simple and easy-to-discriminate features, which makes the classification accuracy of the trained classifier unsatisfactory. [0003] In addition, in the prior art, it is difficult to measure the difficulty of training tasks. Cont...

AI Technical Summary

Problems solved by technology

[0002] In the traditional image classification method, the gradient-based optimization method is used for training. However, the gradient-based optimization training method will reduce the amount of information of the extracted features during processing, so in some cases it will not learn All discriminable features in the image, and only learn those simple and easy-to-discriminate features, which makes the classification accuracy of the trained classifier unsatisfactory

[0003] In addition, in the existing technology, it is difficult to measure the difficulty of the training task

Images