深度神经网络学习的结构基础：自动编码器与限制玻尔兹曼机

[摘要] 自动编码器（AE）和限制玻尔兹曼机（RBM）是在深度神经网络领域广泛使用的两种常见的基础性结构。它们都可以作为无监督学习的框架，通过最小化重构误差，提取系统的重要特征；更重要的是，通过多层的堆叠和逐层的预训练，层叠式自动编码器和深度信念网络都可以在后续监督学习的过程中，帮助整个神经网络更好更快地收敛到最小值点。

[关键词] 深度学习；神经网络；AE；RBM

[Abstract] Auto魛喖encoders (AE) and Restricted Boltzmann Machines (RBM) are two kinds of basic building blocks which are widely used in the architectures of deep neural networks. By minimizing the reconstruction errors, both the AE and the RBM can extract the key characteristics of the input data and can work as the basic framework of the unsupervised learning. Moreover, with the layer-by-layer stacking and layer-wise pre-training, both the stacked AE and the deep belief networks can help neural networks converge faster and better in the following supervised fine-tuning process.

[Keywords] deep learning; neural network; AE; RBM