Intelligent Signal Processing System Design

点击翻译按钮获取中文摘要

An intelligent signal processing system is defined as an intelligent system designed and built to enable applications that can sense to analyze relevant signals in a changing environment, perceive and interpret to model the relevant information from the signals, adapt and learn to process and predict the signals, to imitate, automate, and augment some intelligent behaviors of human beings to serve the needs of the end users. The core technologies to be developed to build intelligent signal processing systems are described below. Recurrent associative memory is a deep recurrent neural network that can learn and recall binary patterns and gray level images. The network allows two modes of memory recall, namely, recalling by a pattern-pair from both the input and the output layers, and recalling by a single-pattern from either the input layer or the output layer. In this part of the research, recurrent associative memory shall be used for feature extraction and/or noise reduction of different types of signals, and fast preprocessing and effective learning algorithms will be developed for efficient and effective operations. Discrete Gabor transform of a real discrete periodic sequence is defined as the discrete Fourier transform of the product of the sequence and a shifted analysis (or Gaussian) window to yield the complex discrete Gabor coefficients for time-frequency analysis. In this part of the research, new theories, methods, and algorithms will be developed for discrete Gabor transform and its multiwindow versions to perform dynamic time-frequency analysis of signals (or images) to effectively extract features. A deep sigmoid processing system consists of a combination of block nonlinear models which can be offline trained to model and process or predict signals as a global model of knowledge memory. The global model is designed to work collaboratively with an online adaptive local model constructed by a block nonlinear model which stores the knowledge memory of the local signals. The combined output of the global and local models obtained from the same input is designed to better process or predict an input signal. The system will be designed to avoid catastrophic forgetting and enable backward transfer learning. In this part of the research, methods and algorithms will be developed to design and build deep sigmoid processing systems with low complexity targeted to day-to-day applications. A deep fuzzy neural system is constructed from a deep sigmoid processing system by replacing the nonlinearity of each block nonlinear model in its global and local models by a fuzzy neuron. The system will be designed to reason and explain conclusions reached from inputs by the deep fuzzy neural system. In this part of the research, methods and algorithms will be developed to design and build deep fuzzy neural systems targeted to more sophisticated and complex applications that require human-like reasoning.