基于二维独立均匀分布对抗自编码的模式识别研究

Pattern Recognition Study of Adversarial Auto-encoder Based on Two-dimen⁃sional Independent Uniform Distribution

标准对抗自编码模型能够以自监督方式自适应提取输入样本的特征,并通过对抗机制对提取的特征施加特定的先验分布,从该先验分布进行采样输入解码器,则可生成与输入样本近似的样本。但在实际应用中,有时需要生成指定类别的样本;对于模式识别任务,通常还需要对多类别样本的特征进行提取,并强化特征间差异,从而进行聚类分析。针对上述需求,本文提出基于二维独立均匀分布对抗自编码的分析模型。在该模型中,根据类别信息构建二维均匀分布,便于对不同类别的可视化特征进行专属约束,从而强化不同类别特征间的差异;此外,通过自监督与对抗训练,使得对应特定类别信息的均匀分布样本能够生成指定类别的样本。方法经网络公开数据MNIST数据集进行了验证,研究表明,该方法能够利用与类别信息相关的二维独立均匀分布对隐变量进行约束,提高了特征聚类性能,并能够生成指定类别的样本。

A standard adversarial auto-encoder(AAE)model can adaptively extract features of the input samples in an unsupervised way,and apply a specific prior distribution to the extracted features through an adversarial mechanism.When samples collected from the prior distribution are fed into the decoder,the real-like samples can be generated.However,in practical applications,it is some⁃times necessary to generate some specific samples.For pattern recognition tasks,features of multi-class samples often have to be ex⁃tracted and enhanced to perform clustering analysis.To satisfy the above needs,an AAE analysis model based on two-dimensional in⁃dependent uniform distribution was proposed.In this model,a two-dimensional uniform distribution was established based on the label information,which allowed to apply specific constraints on the visual features and to enhance the discriminability between features of different classes.In addition,through auto-supervised and adversarial training,specific samples could be generated from the corre⁃sponding uniform distribution samples.Finally,the method was verified by the publicly available MNIST dataset.This study shows that this method can utilize a two-dimensional independent uniform distribution related to the label information to constrain the hidden vari⁃ables and improve clustering performance of the features.The specific samples can also be generated effectively.