基于大脑听觉系统机理的声学事件高层语义的机器认知模式研究

It is one of the main goals in the field of artificial intelligence and information processing that the computer has the ability of human-like auditory system, which is also called machine hearing, and its key problem is the high-level semantic cognition of the acoustic events. The auditory brain system of higher organisms has a gained excellent cognitive ability in the historical evolution process, which has considerable reference significance for recognizing the high-level semantic information of acoustic events. However, most of the current works just following the studies from the areas such as the image processing and machine vision, lack the models and methods that are suitable for characterizing the acoustic events and fail to use the cognitive mechanism of the auditory brain system. And hence, there are still many problems to be solved. In this project, the problem of the high-level semantic recognition of the acoustic events is studied based on the cognitive mechanism of the auditory brain system by using machine learning and optimization method. The main contents studied in the project include the followings: neural response coding in accordance with the auditory cognitive mechanism of cerebral cortex cells and the method of learning the coding model; information abstraction for multiresolution in time-domain; information fusion, semantic abstraction layer by layer, and dimension reduction; high-level semantic selection towards the target. The research work can provide theories and methods for the related application fields, and further promotes the development of the new research field of machine hearing.

计算机人工智能与信息处理研究领域的一个核心目标，是使计算机具有类似人类听觉系统的听觉能力（即机器听觉），其核心问题是声学事件的高层语义认知。高等生物的大脑听觉系统在长期进化过程中获得了出色的声音语义认知能力，这对研究声学事件高层语义认知模式具有重要的借鉴意义。然而，目前这方面的研究大多沿承自图像处理、机器视觉等研究领域，既缺少符合声学事件自身特性的模型和方法，也没有很好地借鉴大脑听觉系统的认知机理，还有许多问题亟待解决。本项目基于大脑听觉皮层的认知机理，以机器学习和优化方法为工具，研究声学事件的高层语义认知模式。项目的主要研究内容包括符合大脑皮层细胞听觉认知机理的神经响应编码方式及其模型学习方法、多分辨率时域信息抽象方法、信息的融合和逐层语义抽象降维方法、面向目标任务的高层语义选择方法等。本项目的研究工作能够为相关应用领域提供理论和方法的支撑，从而推动机器听觉这一新兴研究领域的进一步发展。

本项目重点开展基于大脑听觉机理的声学事件高层语义认知模式研究。在项目的执行过程中严格按照项目计划书要求开展工作，已完成了所有的研究计划内容，并对部分研究内容进行了拓展性研究。在如下几方面取得了重要研究进展：①基于时域和变换域(频域)联合先验的自适应降噪方法、基于Primal-Dual算法的联合域降噪方法；②基于神经细胞群响应的稀疏预测编码方法；③神经细胞群时频接收场学习方法；④时域多分辨率语义编码；⑤基于时频多分辨率时域信息的融合、逐层语义抽象、基于时频局部相似性的时频域特征聚合方法；⑥高层语义认知模式的选择方法。.目前项目组已经按计划任务书的要求基本完成研究内容，提出了若干具有自主知识产权的方法。在本领域高水平期刊和会议上发表论文10篇，其中SCI收录的论文6篇，4篇重要会议论文，10篇论文全部是EI检索源。申请发明专利2项，其中1项已获授权并进行了成果转化应用。培养硕士研究生9名。