Robot Brain Project CREST Development of Brain-Informatics Machines through Dynamical Connection of Autonomous Motion Primitives
Robot Brain Project
Nakamura Group
Asada Group
Tsuchiya Group
Ushio Group
Yoshizawa Group
Sasaki Group

Results | Yoshizawa Group

Modeling Visual Recognition and Online Linear Discriminant Analysis
Shuji Yoshizawa*1,Hiroshi Mizoguchi*2
*1Saitama University,*2Tokyo University of Science

Aiming development of brain-oriented recognition and decision technologies, our group has studied three subjects.

The first subject is modeling of brain information processing; its goal is understanding of information processing mechanism in the brain with neuron level and neural network level. We have clarify the mechanisms of synaptic regulation and proposed multiplicative spike-timing-dependent synaptic plasticity (STDP) rules which can achieve synaptic regulation. We have also investigated the property of the connection which is required for realization of the qualitative results shown by the psycho-physical experiments on the binocular rivalry.

The second subject is the study on recognition technology. Only from a two-dimensional movie, we can recognize three dimensional positions and motions of a person in it. Though the mechanism of that ability is not known completely, we have a conjecture that a model of human body is built in the brain and it is applied to two-dimensional images so that natural motions can be estimated. According to this strategy, we have tried estimation of three-dimensional motions based on positions of characteristic body points which are detected from two-dimensional movies. In experiments on qualitative properties for the case that noises are added to the detected positions, an interesting behavior is observed; unless the noises are so large that no candidate of three-dimensional posture remains, naturalness of estimated motion has little dependence on the amount of noises.

The third subject is the study on discrimination technology. The brain can learn new things and it has adaptability to change of situations during the learning. It can also extract useful components for its decision from high-dimensional input data. In order to develop discrimination technologies which have the above nature, we have constructed online linear discriminant analysis (OLDA) and related techniques. In addition to theoretical investigation of them, we have also built several vision systems as applications of them, face recognition system, arbitrary pattern recognition system, human walking pitch extraction system based on visual tracking of person's heel, and so on.

Yoshizawa Group
PDF(2.33MB) ]