基于多粒度认知的智能计算研究

在传统的机器学习研究中，数据空间与知识空间分离表达、计算机的信息处理过程与人脑的认知处理过程不一致，成为了当前人工智能研究需要解决的核心关键问题。本文从认知计算的角度，回顾分析了基于多粒度认知的智能计算研究的发展历史轨迹，介绍了该领域的研究现状，提出了多粒度认知计算、可解释的认知机器学习、脑认知的智能计算辅助等该领域的三个前沿研究方向，探讨了在多粒度认知启发下，这些智能计算研究的未来可能发展趋势。

Granular computing (GrC) is a machine intelligence and cognitive computing methodology that uses granule as a processing object. It is a powerful tool for approximate solution of complex problems at multiple levels and scales. Its essence is to simulate the multi-granularity cognition mechanism of the human brain, and establish a set of theories and methods for information space transformation between the information processing mechanism of computer and the multi-granularity cognition process of the human brain. From the perspective of granular computing and cognitive computing, this paper analyzes several contradictory phenomena and problems existing in artificial intelligence research. The information processing mechanism of computer starts from samples on fine granule. It extracts knowledge from data based on the expression of data space. Nevertheless, the human cognition process maps and reasons between data and knowledge based on the expression of knowledge space. In traditional machine learning research, there exists the problem of separate expression of data space and knowledge space. The separate expression leads to the independence of data and knowledge. It is difficult to establish the mapping and reasoning from data to knowledge. In the process of image recognition, the computer algorithm processes from the pixel points rather than high-level semantic features or concepts of the image. However, the human visual cognition process starts with global topological features, and then gradually refined features. This contradictory phenomenon shows that the information processing mechanism of computer is not consistent with the cognition process of the human brain. This leads to a number of serious problems, such as the vulnerability and lack of interpretability of deep learning neural network models. For example, adding a small amount of specific noise to an image may seriously reduce the recognition performance of a deep neural network and generate a completely wrong recognition result. In addition, deep convolutional neural network models trained by ImageNet tend to rely more on fine-granular features such as texture for recognition (texture bias) than coarse-granular features such as shape. In this paper, the historical development trajectory of multi-granularity cognition based intelligent computing (MGCIC) is reviewed and the current research progress about the nine major issues of data-driven granular cognitive computing (DGCC) is introduced. Furthermore, the three frontier research directions such as multi-granularity cognitive computing, interpretable cognitive machine learning, and intelligent computing assistance for brain cognition are proposed. In detail, multi-granularity cognitive computing is based on granular computing, which integrates the problem-solving mechanism of the human brain into granular computing models. It integrates the cognition process of the human brain and the information processing mechanism of computers. Through collaboratively inspecting the knowledge implied in a multi-granularity knowledge space, these two seemingly opposite transformation processes could be studied in the same multi-granularity space. The objective of interpretable cognitive machine learning is to find out the complex correlation between concepts distributed in a deep knowledge space generated from massive data and to establish the mapping between a deep knowledge expression space and a deep neural network feature space. The objective of intelligent computing assistance for brain cognition is to promote the research of brain science and cognition science with the help of intelligent computing models and methods based on the idea of computer simulation. This paper provides new ideas and inspirations for the intersection and fusion study of artificial intelligence and human (cognitive) intelligence.