auditory feature extraction and binaural processing mechanisms follow information theory withsparse representation. The ICA-based features resemble frequency-limited features extractedfrom the cochlea and also more complex time-frequency features from the inferior colliculus andauditory cortex. The top-down attention model shows how the pre-acquired knowledge in ourbrain filters out irrelevant features or fills in missing features in the sensory data. Both thetop-down attention and bottom-up binaural processing are combined into a single system forhigh-noisy cases. This auditory model requires extensive computing, and several VLSIimplementations had been developed for real-time applications. Experimental resultsdemonstrate much better recognition performance in real-world noisy environments.

PL9: VISUAL PERCEPTUAL LEARNINGZhongzhi Shi, ProfessorInstitute of Computing Technology, Chinese Academy of Sciences, CHINA

ABSTRACTPerceptual learning should be considered as an active process that embeds particular abstraction,reformulation and approximation within the Abstraction framework. The active process refers tothe fact that the search for a correct data representation is performed through several steps. A keypoint is that perceptual learning focuses on low-level abstraction mechanism instead of trying torely on more complex algorithm. In fact, from the machine learning viewpoint, Perceptuallearning can be seen as a particular abstraction that may help to simplify complex problemthanks to a computable representation. Indeed, the baseline of Abstraction, i.e. choosing therelevant data to ease the learning task, is that many problems in machine learning cannot besolve because of the complexity of the representation and is not related to the learning algorithm,which is referred to as the phase transition problem. Within the Abstraction framework, we usethe term perceptual learning to refer to specific learning task that rely on iterative representationchanges and that deals with real-world data which human can perceive.In this talk we focus on sparse coding theory and granular computing model for visual perceptuallearning. We propose an attention-guided sparse coding model. This model includes two modules:nonuniform sampling module simulating the process of retina and data-driven attention modulebased on the response saliency. Based on tolerance relation we construct a more uniformgranulation model, which is established on both consecutive space and discrete attribute space.

PL10: QUOTIENT SPACE MODEL BASED HIERARCHICAL MACHINELEARNINGLing Zhangt , Professor, and Bo Zhang2, Professor'Anhui University, CHINA2Tsinghua University, CHINA, Member of the Chinese Academy of Sciences

ABSTRACTWe proposed a quotient space based model that can represent the world at different granularitiesand can be used to handle problems hierarchically. The model can be used in two different ways:top-down deduction and bottom-up induction. In this paper, we will discuss the quotient spacemodel based bottom-up induction, i.e., hierarchical learning. Some approaches for learning thestructural knowledge from data are presented. The main advantage of hierarchical induction is itsefficiency, that is, the whole structure of data can be abstracted at once.

xiv