Ryuichi Oka, President of the University of Aizu, Japan
Title: Field Computation Architecture of Algorithms for Intelligent and Smart Pattern Processing
Abstract: In my talk, we propose a scope called field computation architecture as a scheme of hardware implementation. In the real world, there are two types of information. One is pattern information such as time signal, two-dimensional image, video composed of time and the two-dimensional image. Pattern information has its own intrinsic neighborhood relation between elements. Another one is symbol information such as natural language, the network of words, list of symbols, etc. Each element of symbol information has its own intelligent meaning rather than an element of pattern information. We consider in this talk only the processing of pattern information. In the field of pattern processing, field computation is frequently used for extracting features of pattern like edge, line, corner, filtering using a mask, etc. However, the conventional field computation provides microscopic or low intelligent entity in most of the cases. These field computations are regarded as a kind of pre-processing of the pattern. Then main algorithms including symbol processing are devised for final intelligent results by using the results of pre-processing. We focus on field computation algorithms realizing intelligent and smart pattern processing. This means that the field computation can provide intelligent and smart results which are used as final results by aggregating of all local elements, then we don’t need further processing for final results. Intelligent and smart pattern processing uses an array of state each element of the array corresponds to a working memory as well as pre-processing. The difference is that each working memory can store both microscopic and macroscopic information existing in the whole pattern, and also the obtained information of each memory is location dependent. The location-dependent intelligence can work for segmentation-free recognition of the pattern. Because segmentation of pattern is a procedure conducted in a location-dependent way which is based on suitable and optimal aggregation of local information. Location-dependent segmentation can be done by evaluating each state of the array. How to aggregate local information is a key feature of the algorithm. Field computation is a local and parallel and iterative processing for aggregating local information to make macroscopic intelligence of the pattern. Creation of algorithm provides a new computational framework which requires a unique hardware implementation because a new algorithm is developed without consideration of existing hardware architecture. The progress of hardware needs to meet promising new architecture provided by intelligent and smart algorithms. We want to use hardware without additional and complex software for obtaining final results. In this talk, we show five algorithms and their smart functions and experimental results as references for considering new hardware architecture. The five are: 1) Continuous Dynamic Programming for optimal, non-linear, and segmentation-free recognition of a reference in an endless stream of input sequences, 2) Two-dimensional Continuous Dynamic Programming used for a) segmentation-free recognition of image, and b) segmentation-free and speaker-independent recognition of a single speech spoken by multiple speakers, 3) Time-space Continuous Dynamic Programming for time-space segmentation-free recognition of air-drawn gesture and handwritten characters, sport performance (figure skating, football, etc.), motion tracking and flow line extraction from video capturing a running or walking cloud of people, and cars, etc. 4) Accumulated-Motion-Parallax Method for 3D image reconstructing of a wide scene of a city, indoor/outdoor scene from a piece of video. 5) an algorithm for reconstructing the dynamic 3D scene from a video captured by an onboard camera on a car. The dynamic 3D scene can be used as a distance sensor for automatic car driving without laser device.
Biography: Dr. Oka graduated from a master course in mathematical engineering. University of Tokyo, Japan, March 1970 and worked for Electrotechnical Laboratory of Ministry and Industry and Trade from April 1970 as a researcher and received Dr. degree of engineering (1984) from The University of Tokyo. He stayed at National Research Council of Canada as a visiting scientist (July 1984 — June 1985). He worked for a ten-year national project of Japan called Real World Computing (March 1993 – March 2002) as a director of the division and a chief of laboratory. His research areas include character recognition, speech recognition and retrieval, image retrieval, image processing, image understanding, computer vision (3D scene reconstruction from a video), data mining, data visualization, and mobile robot, drone network connected by cables, etc. He proposed a family of so-called Continuous Dynamic Programming for segmentation-free optimal matching between two sequence patterns, two two-dimensional patterns, two three dimensional patterns and two time-space patterns. These matching methods are widely used for researchers and making industrial products.
Keynote II: TBC
Keynote III: TBC