基于语义自动编码机的零次学习研究

Research on Semantic Auto-encoder for Zero-Shot Learning

零次学习是计算机识别中的一个重要研究领域,具有广阔的应用前景和潜力。零次学习致力于从已知数据中提取有效特征,以用于对未知数据进行分类或识别。自动编码机将原空间特征转化为编码空间特征,并允许逆向变换,可以一定程度保留原始特征空间的分布。在传统自动编码机的定义上加入限制,使其编码层可以兼容语义特征,使编码过程无需迭代求解。论文分别通过加入正则项降低模型的过拟合性,以及通过核函数进行改进的两种思路入手,使最终效果得到了提升。实验结果达到了目前先进水平。

Zero-Shot Learning(ZSL)is one of the most important research fields for computer science with broad application prospects and potentials.Zero-Shot Learning algorithms recognize or classify test samples through features extracted from training samples,and any test samples are forbidden during the training phrase.Auto-encoder projects the original features into certain code space and inverse operation is offered simultaneously which makes this structure reserving the distribution of the original feature space.By adding some constraints,it makes auto-encoder compatible with semantic features.The answer can be obtained by solv ing a Sylvester equation,w ith regularization or kernel trick for developing its behavior.Experimental results achieve the leading lev el in present.