在各种研究工作中，已经探索了自动编码器的结合，以提高分类器的性能。自动编码器表现出一系列的变化，每个都针对特定的目标，如去噪，特征提取，生成合成数据和降低维度。部署自编码器的核心原则是在输出端复制输入数据，同时保持最大相似度。编码器和解码器都由一个编码器和一个解码器组成，一旦编码器被训练，它与分类器的集成无缝地增强了结果。值得注意的是，在具体的研究中，自动编码器被用于消除噪声和增强结果[ 24 ]。

% 自编码器代替PCA 进行无监督特征学习 . 自编码器的变体 将自动特征学习和降维结合到一个集成神经网络中进行活动识别
Autoencoders are used to replace Principal Component Analysis (PCA) for unsupervised feature learning.
Variants of autoencoders integrate automatic feature learning and dimensionality reduction into a single neural network for activity recognition.

优点

  • 自编码器是一种前馈神经网络，它学习输入样本 (通常通过引入瓶颈层来实现)的紧凑潜在表示，并被训练以重建输入样本。与基于线性变换的方法相比，基于自编码器的模型具有优势，因为自编码器能够学习到最能表征数据的流形结构。此外，这种方法还可以通过最小化一些差异度量来显式地减少领域差异。

缺点

  • VAE-based semi-supervised methods is that the latent representations are stochastically sampled from the prior distribution instead of being directly rendered from the explicit observations.

变种

SDAE

  • The stacked denoising autoencoders learns features in an unsupervised manner by using the greedy layer-wise training method
  • The training process of the stacked denoising autoencoders starts by training the first layer on the input to learn the features in the first hidden layer. Similarly, this operation is repeatedly conducted for subsequent layers to finish the whole training process. After the layer-wise training, a finetuning operation is followed to optimize the parameters of all layers through backpropagation, which will further improve the results.
  • Masking Noise Level: The masking noise level indicates the fraction of elements of the input data that are forced to 0. Fig. 6 (a) demonstrates the influence of different masking noise levels on the classification accuracy. The best performance arises at around the corruption ratio (or masking noise level) of 0.3, from which increasing or decreasing the ratio will lead to a decrease in the classification accuracy. Note that when the corruption ratio equals 0, the model becomes a stacked sparse autoencode

VAE

  • VAE is a variant of autoencoder that is able to denoise motion signals. As autoencoder, VAE encodes the input data into a feature-representation vector in the latent space. However, in VAE, the feature representation vector is constrained to Gaussian distributions [33]. Also, VAE uses two loss functions, a combination of Kullback-Leibler divergence (KL) and Mean Square Error (MSE). The first loss function forces the feature-representation vector in the latent space to follow a Gaussian distribution, measuring the relative entropy between the approximate posterior and the prior probability density function [33]. The second loss measures the similarity between the output and the input