CAJ | 학술논문

受限波尔兹曼机(restricted Boltzmann machines, RBM)是一类具有两层结构、对称连接且无自反馈的随机神经网络模型，层间全连接，层内无连接。近年来，随着RBM的快速学习算法—对比散度的出现，机器学习界掀起了研究RBM理论及应用的热潮。实践表明，RBM是一种有效的特征提取方法，用于初始化前馈神经网络可明显提高泛化能力，堆叠多个RBM组成的深度信念网络能提取更抽象的特征。鉴于RBM的优点及其在深度学习中的广泛应用，本文对RBM的基本模型、学习算法、参数设置、评估方法、变形算法等进行了详细介绍，最后探讨了RBM在未来值得研究的方向。
수한파이자만궤(restricted Boltzmann machines, RBM)시일류구유량층결구、대칭련접차무자반궤적수궤신경망락모형，층간전련접，층내무련접。근년래，수착RBM적쾌속학습산법—대비산도적출현，궤기학습계흔기료연구RBM이론급응용적열조。실천표명，RBM시일충유효적특정제취방법，용우초시화전궤신경망락가명현제고범화능력，퇴첩다개RBM조성적심도신념망락능제취경추상적특정。감우RBM적우점급기재심도학습중적엄범응용，본문대RBM적기본모형、학습산법、삼수설치、평고방법、변형산법등진행료상세개소，최후탐토료RBM재미래치득연구적방향。
A restricted Boltzmann machine (RBM) is a particular type of random neural net-work model which has two-layer architecture, symmetric connections and no self-feedback. The two layers in an RBM are fully connected but there are no connections within the same layer. Recently, with the advent of a fast learning algorithm for RBMs (i.e., contrastive divergence), the machine learning community set off a surge to study the theory and applications of RBMs since it has many advantages. For example, a RBM provides us an effective tool to detect features. When a feed-forward neural network is initialized with an RBM, its generalization capability can be significantly improved. A deep belief network composed of several RBMs can detect more abstract features. Due to the advantages and wide applications of RBMs in deep learning, this paper attempts to provide a introductory guide for novice. It presents a detailed introduction of basic RBM model, its representative learning algorithm, parametric settings, evaluation methods, its variants and etc. Finally, some research directions of RBMs that are deserved to be further studied are discussed.