Incremental learning method based on generative adversarial network knowledge distillation

Abstract

The invention relates to an incremental learning method based on generative adversarial network knowledge distillation. The method comprises the following steps: S1, enabling a generator to generate an old pseudo training sample with a sample label according to input condition information, and combining the old pseudo training sample with a conventional new training sample, thereby forming a new training data set; S2, training a new generator and a discriminator on the original generative adversarial network according to the new training data set; S3, training a classifier based on the new training data set, training the classification capability of the classifier by utilizing the original label information of the sample in the training process, and guiding the learning of the classifier by taking a new discriminator as a teacher model to obtain a new classifier for generating a high-quality image. According to the method, the total category related information of the new category andthe old category is effectively provided for the classification model, so that forgetting of the classification model in incremental learning can be reduced, and classification precision is improved.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, and relates to an incremental learning method based on generative confrontation network knowledge distillation. Background technique [0002] A deep learning network is a machine learning method that mimics the learning process of the human brain. At present, great progress has been made in digital image processing and natural language processing. However, unlike the human learning process, the current mainstream deep learning network algorithms all aim at single task learning. Human learning is a process in which the acquired knowledge is continuously incrementally accumulated with the deepening of learning. Incremental learning is a topic set up to study the deep learning network to simulate this learning process. It is of great and far-reaching significance to the development of deep network technology. [0003] Among the current mainstream incremental learning methods, a r...

AI Technical Summary

Problems solved by technology

The work of Rebuffi et al. is to select some training samples from old tasks for data reproduction, but this method needs to store samples. The samples of the training set are sometimes limited by patents or privacy rights and other reasons and cannot be published, and storing samples requires additional the cost of

Images