先进储备池神经计算方法及其在时间模式识别中的应用

Temporal pattern recognition is one of the key and common issues for intelligent systems such as video tracking, speech recognition, which require the algorithms should be with temporal order modeling capability. Reservoir computing (RC) is a novel kind of recurrent neural network (RNN) method. It explores the separate "reservoir" and "readout" architecture, and makes the difficult RNN training problem an easy linear regression or classification problem. It has been demonstrated that RC can provide competitive solutions to temporal pattern recognition. However, there are still lots of theoretical and practical issues on the RC requiring to be further improved, such as generalization, adaptation. Therefore, concentrating on the temporal pattern recognition problems, and taking one kind of automatic speech recognition problems for example, this project studies RC generalization performance based on the structural risk minimization principle, and also studies the adaptation issue based on parameter self-adaptation and optimization techniques. Finally, this project will propose a kind of adaptive and structural-risk-minimization compatible algorithms, and also realize an automatic in-vehicle speech recognition prototype system. The validity and superiority of the proposed algorithms are approved by the system. The successful implementation of this project helps to the progresses of neural network theoretical and applied research, besides provides more efficient ideas and solutions for real temporal pattern recognition.

时间模式识别是视频跟踪、语音识别等智能系统的关键共性问题之一，要求算法具备时序建模能力。储备池神经网络是一类新型递归神经网络（RNN），采用"储备池"+"读取器"网络结构组织，促使棘手的RNN训练问题转化为一个简单的线性回归或分类问题。现有研究表明，储备池神经计算方法可为时间模式识别问题提供有竞争力的方案，但在泛化性能、问题适应性等方面还存在大量的理论与实际问题亟待解决。综上，本项目面向时间模式识别问题，以一类自动语音识别问题为例，结合结构风险最小化原理方法，开展储备池神经网络泛化能力研究；结合参数自适应与优化技术，开展储备池适应性研究。最终将建立一种自适应的结构风险最小化储备池算法，并实现一个车载自动语音识别原型系统，实际验证算法的有效性与先进性。本项目的成功实施有助于神经网络理论与应用研究发展，并为求解实际的时间模式识别问题提供新的思路与方法。

本项目面向时间模式识别与分类问题，开展神经计算相关的机器学习方法与应用研究。①针对语音识别问题：提出了一种基于线性降维和受限制玻尔兹曼机（RBM）的孤立数字语音识别方法，针对TI46孤立数字语音库和NOISEX-92噪声数据集的实验表明，该方法可以达到96.1%的识别率，并且可获得显著改善的噪声鲁棒性。提出并实现了一种语音指令控制的智能型电动玩具小车原型开发。②针对时间序列预测与突变检测问题：提出了一种改进的进化灰色模型算法，用于竞争性电力市场短期电价时间序列预测，获得改善的预测精度。提出了一种基于回响状态神经网络（ESN）的非静态时间序列突变检测方法，该方法本质上是将非静态时间序列突变检测问题视为一类分类问题处理。③针对传感器网络中目标跟踪问题：基于条件后验克拉美–罗下界(CPCRLB)提出一种传感器网络节点管理方法。对具有随机传感器故障的多传感器融合估计问题，提出了一种传感器故障容错的多传感器目标跟踪算法。上述研究成果可为求解孤立数字语音识别、时间序列突变检测与预测、多传感器网络目标跟踪等典型问题提供一定的思路和先进算法支持。