Week 9: 深度学习补遗：自编码器与生成式模型

Week 9: 深度学习补遗：自编码器与生成式模型

摘要

本周，跟随李宏毅老师的课程学习了自编码器和生成式模型的课程，两方面的内容关系比较紧密。主要从抽象角度以及一些数学角度对自编码器进行了学习，编码器-解码器架构作为现在模型的一大主流结构，具有比较重要的学习意义，对自编码器以及生成式模型的学习使我对编码器-解码器架构有了一定了解。

Abstract

This week, I followed Mr. Hung-yi Lee’s course to learn about self-encoders and generative models, which are closely related to each other. Mainly from an abstract point of view as well as some mathematical point of view on the self-encoder learning, encoder-decoder architecture as a major mainstream structure of the model now, has a more important learning significance, on the self-encoder as well as the generative model of the learning of the encoder-decoder architecture so that I have a certain understanding of the encoder-decoder architecture.

1. t分布随机邻居嵌入 t-Distributed Stochastic Neighbor Embedding

LLE与拉普拉斯映射只考虑了“临近点必须接近”，并没有考虑“非临近点要远离”。所以实际上，LLE在会把同一个类别的点聚在一起的同时，难以避免不同的类别的点重叠在一起。导致类别之间难以区分。

而t-SNE的要点是，在原空间内计算每一对点$x^i$与$x^j$之间的相似度$S(x^i,x^j)$，然后做归一化$P(x^j|x^i)=\frac{S(x^i,x^j)}{{\sum }_{x\ne k}S(x^i,x^k)}$。在降维之后获取到的点$z^i$、 z j z^j z