RWC 2000 Technical Session

1. Real World Intelligence

1.1 Overview

The 2000 RWC Symposium took place with only two years remaining for the RWC Project. The project is characterized by many research themes which are already at the stage of implementing intermediate results before completion, or will soon be implemented after the prototyping or experimental stage. In Real World Intelligence, there have been promising results such as CrossMediator, which has already been implemented in products. With this background, many laboratories presented their results at the session using digital posters.

At the core of the real world intelligence research lie information integration and learning. The aim is to develop fundamental technologies to add real world intelligence to the conventional information technologies which can accept real world information as they are, recognize or estimate the environment or situations, and autonomously respond to them by integrating multimodal information such as speech, (still/motion) images, and text, which have been handled separately in the past.

Specifically, our research involves the following five domains:
(1) Multi-Modal Functions: Development of agent-driven human interface to use information systems combining speech and images.
(2) Autonomous Learning Functions: Development of agent systems which can autonomously move around in real environments and take proper actions by gathering and learning the environment and information surrounding them through sensing or interactions.
(3) Self-Organizing Information-Base Functions: Development of agent systems which can consolidate, summarize, retrieve, and present a variety of information in the real world or on information networks in a self-organizing manner.
(4) Theoretical and Algorithmic Foundation: Establishment of foundation for information integration and learning/self-organization technologies.
(5) Real-World Adaptive Devices: Development of real-time adaptive devices with adaptability at the hardware level for achieving information integration and learning.
The RWC Project is the first to tackle information integration and yield results, and will surely have a great impact on information processing in the future.

The results are described in the following pages, using CrossMediator (Fig. 1) as a typical example of the results already implemented.

The conventional databases are mainly relational ones, which basically use text for information retrieval. Data may include images and/or speech, but retrieval is done through text. In contrast, CrossMediator, in which each data item includes multimodal information such as speech, still image, motion image, and text, enables us to search multimodal information using any type of information as the key. Systems which can search for a motion image that contains speech used as a keyword are already available as products, and have attracted interest from many sectors. They allow users to easily retrieve the desired information from enormous amounts of motion images without manually tagging motion images. Other developments include ASKS, which handles text information only, and an entry system for PCs using gestures and speech.

There are many other technologies that have already been implemented or soon will be implemented, including static adaptive devices, large-scale simulation system FTA, sign language recognition system, and visualization of many images as search results.

(Yoshikuni Okada, Manager, Research Planning Department)

1.2 Listing of Research Themes in Real World Intelligence

RWCP

Domain	Research Theme	Laboratory	DP
Multi-Modal Functions	Multi-Modal Interface System for the CrossMediator	Multi-Modal Functions TRC Lab	64
	Robust Face and Facial Expression Recognition	Multi-Modal Functions KRDL Lab	64
	Hand Motion Estimation based on a Generic Model	Multi-Modal Functions Sanyo Lab	67
	Multimodal Agent Interface for Communication	Multi-Modal Functions Sharp Lab	66
	Image Processing Camera Module for Video Surveillance	Multi-Modal Functions Mitsubnishi Lab	-
	Sign Language Recognition and its Applications	Multi-Modal Functions Hitachi Lab	65
	Where in the Brain is Responsible for Character Recognition ?	Multi-Modal Functions NTT Lab	67
Information Base Functions	The CrossMediator -An integrated Multi-media retrieval system -	Information-Base Functions TRC Lab	64
	Semantic Information Retrieval from Tagged Text Corpus	Information-Base Functions Mitsubishi Lab	67
	Toward semantics level indexing and retrieval of images and video	Information-Base Functions KRDL Lab	65
	Friendly Information Retrieval through Adaptive Restructuring of Infromation Space	Information-Base Functions TOshiba Lab	-
	VIsualization for Similarity-Based Image Retrieval Systems	Information-Base Functions Hitachi Lab.	65
Autonomous Learning Functions	Situation Decomposition Algorithm for Autonomous Learning	Autonomous Learning Functions Fujitsu Lab.	68
	Feature selection for appearance-based robot localization	Autonomous Learning Functions SNN Lab.	-
	Difficulties when applying Learning and Adaptation to Reconfiguration and Mating Problems	Autonomous Learning SICS Lab.	-
	Non-speech sound recognition with microphone array	Autonomous Learning MRI Lab.	68
Adaptive Device	The Design and Implementation of an ALUBased Reconfigurable Adaptive Device	Adaptive Device NEC Lab.	69
	8庁ﾟ8 Digital Smart Pixel Array	Adaptive Device Matsushita Lab.	68
Theoretical Foundation	Methodology of Distributed and Adaptive Learning	Theoretical Foundation NEC Lab.	-
	Logical Reasoning of Patterns -A New Solution for Knowledge Acquisition Problem -	Theoretical Foundation Toshiba Lab.	-
	Artificial Reality : Mapping Real -World Processes into Fine-Grained Multiagent Systems	Theoretical Foundation GMD Lab.	69
	Variational methods for approximate reasoning in graphical models	Theoretical Foundation SNN Lab.	69
	Stochastic Pattern Computing	Theoretical Foundation SICS Lab.	-

Electrotechnical Laboratory(ETL) RWI Center

Research Theme	LAboratory	DP
Inference and Learning with Graphical Models	ETL-RWI Center, Learning and Information Integration Lab.	-
Adaptive Vision Systems for Real-World Intelligence -A Wearable Vision System-	ETL-RWI Center,Adaptive Vision Lab.	66
A Real-time Filled Pause Detection System - Toward Spontaneous Speech Dialogue-	ETL-RWI Center,Interactive Intermodal Integration Lab.	-
Design and Application of Multimodal Common Format	ETL-RWI Center, Language Integration Lab.	-
Integration of Real-world Interaction FUnctions on the Jijo-2 Office Robot	ETL-RWI Center,Jijo-2 Robot Lab.	-
Dynamic Adaptive Devices and their Applications	ETL-RWI Center, Evolvable System Lab.	66
Large-scale Optical Neural Network with High-speed Learning	ETL-RWI Center, Information Optical Lab.	-
Intergrating Framework for Application Programs with Network Services Using a Web Browser	ETL-RWI Center, RWC Library Lab.	-

2. Parallel and Distributed Computing

2.1 Overview

Seamless parallel and distributed computing in this context means: (1) an environment for users to use a parallel and distributed system as a single computer, (2) parallel and distributed computers connected via network infrastructure such as a high-speed LAN, and (3) a heterogeneous computing environment.

SCore Cluster System Software can be used in a homogeneous computing environment in the first stage and in a heterogeneous computing environment in the second stage. It allows us to build a parallel system under UNIX-based operating systems (i.e., Linux, NetBSD, Solaris). Therefore, we are able to realize a so-called PC cluster with Linux on PCs, and develop cost-effective high-performance computers which allow supercomputing and large-scale data mining. The software is available on media such as WWW and CD-ROM, and is already used for applications such as simulations by many research institutes, including University of Bonn (Germany) and Lyon University (France) in Europe, and Los Alamos National Laboratory in the USA. Back in Japan, Mitsubishi Electric Corporation uses it for real-time simulation of power systems thanks to its fast communication speed. The software is expected to play an important role in Linux-based PC clusters.

Regarding languages, WPP, which converts sequential programs into OpenMP by automatically parallelizing them, and OmniOpen MP, an OpenMP compiler, have been developed, as well as RHiNET II which boasts the world's fastest communication switching speed. Using these, a machine will be constructed to demonstrate seamless parallel computing.

(Yoshikuni Okada, Manager, Research Planning Department)

2.2 Lists Tof Research Themes in Parallel and DIstributed Computinge