Abstract: Virtual sample generation technology mainly makes research on how to combine the priori knowledge of a domain with the existing training samples to create additional training samples, enlarging the training dataset, and improving the generalization ability of learning machine. As a method of bringing priori knowledge to machine learning, virtual sample generation technology has become a primary means to improve the generalization ability on small sample learning problem, and has been studied widely by scholars home and abroad. In this paper, firstly the concept of virtual sample was introduced, two indexes of evaluating the capability of virtual sample generation technology were given, and the influence of virtual sample generation technology on the generalization ability of learning machine was discussed.Then the current virtual sample generation technologies were divided into 3 classes by their essence, and for each class several typical virtual sample generation technologies were discussed. Moreover the deficiencies of the current virtual sample generation technology were pointed out Finally, made a summary and proposed our own viewpoint on the further development of virtual sample generation technology.

Key words: Virtual samples, Prior knowledge, Generalization ability, Machine learning, Small sample learning problem