Architecture and learning of multilayer convolutional adaline modules

• Main Authors: Yen-Chih Lu, 盧彥志
• Other Authors: Jiann-Ming Wu
• Format: Others
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/79386022005368803239

碩士 === 國立東華大學 === 應用數學系 === 103 === Abstract This work proposes multilayer convolutional adaline (adaptive linear element) modules for signal and time series processes. A convolutional adaline is extended to be composed of a linear receptive field, a U-shape transfer function, and an integrator. The input layer is equipped with a set of temporal delays for auto-regressive sampling of segments from a time series. The hidden part contains a receptive field and a U-shape transfer function. The output of a convolutional adaline is an integrator that cumulates square auto-regressive approximating errors over a temporal period. A module of multilayer convolutional adalines has two operation modes, a deterministic activation mode and a stochastic activation mode, respectively generating a winner-take-all response and the expectation of a discrete response. The architecture of multiple convolutional adaline modules is proposed in this work. The receptive fields in a convolutional adaline module are derived by unsupervised learning subject to many time series. The goal is to seek best linear filters for time series clustering. The posterior weights of translating module activations to the visible output are refined by supervised learning for time series identification. The propose supervised and unsupervised learning approaches are shown reliable and effective for constructing multilayer convolutional adaline modules. Keywords：adaptive linear element, convolutional adaline, linear receptive field, U-shape transfer function, autoregressive approximation, sound clustering, time series identification, sound recognition, multilayer neural networks