Small sample learning algorithm based on covariance measurement

Abstract

The invention discloses a small sample learning algorithm based on covariance measurement, and belongs to the field of computer vision. The algorithm comprises the steps: (1) introducing an interpolation training mechanism to learn migratable knowledge; (2) designing a local covariance representation, and then embedding the local covariance representation into a deep network to learn and express each concept; (3) constructing a covariance measurement layer to measure the distribution consistency between a query sample and the concepts based on the local covariance representation. According tothe method, a novel and concise end-to-end covariance measurement network CovaMNet is provided, a second-order local covariance representation is designed to replace traditional first-order concept representation, and a new covariance measurement function is provided. A comparison experiment result on a plurality of reference data sets is analyzed to obtain that the CovaMNet framework provided bythe invention shows a competitive effect on a general small sample classification task and a fine-grained small sample classification task.

Description

technical field [0001] The invention relates to a small-sample learning algorithm based on covariance measurement, which belongs to the field of computer vision. Background technique [0002] Humans can learn new concepts from very few examples and have strong generalization capabilities, which are not currently available in machine learning algorithms, that is, humans can learn a new concept from one or a few examples, But standard machine learning algorithms require many more instances to barely achieve the same capability. At present, machine learning algorithms rely too much on labeled data, and in practical applications, the cost of data labeling is often high. How to use a limited amount of labeled data for learning and make the model have a high generalization ability has become an important issue. [0003] For the above problems, it is necessary to utilize more prior knowledge to help representation and learning. [0004] Existing research programs such as figure ...

AI Technical Summary

Problems solved by technology

Kelsey R.Allen et al. proposed to use infinite mixture prototypes to represent each category based on the prototype network. In the prototype network, each category is only represented by a single prototype, which is difficult to fully express the complex distribution of categories.

Images