Multi-sample AutoMix data enhancement method and system based on adversarial learning

• Main Authors: LIAO HONGCHAO, YANG KAIYUAN, JIN XIN, QIN HUAFENG, LI HAIYANG, ZHU HONGYU
• Format: Patent
• Language: Chinese; English
• Published: 12.04.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention discloses a multi-sample AutoMix data enhancement method and system based on adversarial learning. The method comprises the following steps: acquiring a sample data set S and a corresponding label; a sample data set is subjected to data enhancement through an AdAutoMix model, the AdAutoMix model comprises a mixed sample generation module and a sample classification module, the mixed sample generation module comprises an encoder and a mixer, the sample classification module comprises a classifier, the mixed sample generation module randomly selects N samples from the sample data set S to form a multi-sample set, and the multi-sample set comprises N samples; and for each multi-sample set X, a mixed sample xmix is generated based on an attention mechanism, and the mixed sample generation module and the sample classification module are jointly optimized through adversarial learning. According to the method, more diversified samples can be generated, and the generalization ability of the model is imp