《Deep Learning and Its Applications to Signal and Information Processing 》.pdfVIP

[exploratory DSP] Dong Yu and Li Deng Deep Learning and Its Applications to Signal and Information Processing oday, signal processing INTRODUCTION TO DEEP LEARNING each lower layer’s outputs are fed to its research has a significantly Many traditional machine learning and immediate higher layer as the input. The widened its scope compared signal processing techniques exploit shal- successful deep learning techniques with just a few years ago [4], low architectures, which contain a single developed so far share two additional key and machine learning has layer of nonlinear feature transformation. properties: the generative nature of the T been an important technical area of the Examples of shallow architectures are model, which typically requires an addi- signal processing society. Since 2006, conventional hidden Markov models tional top layer to perform the discrimi- deep learning—a new area of machine (HMMs), linear or nonlinear dynamical native task, and an unsupervised learning research—has emerged [7], systems, conditional random fields pretraining step that makes effective use impacting a wide range of signal and (CRFs), maximum entropy (MaxEnt) of large amounts of unlabeled training information processing work within the models, support vector machines (SVMs), data for extracting structures and regular- traditional and the new, widened scopes. kernel regression, and multilay