Primordial Knowledge Model Core
Summary ・ Theme ・ Member ・ Contact
■All Themes
The Primordial Knowledge Model Core -- Research and Education
Primordial knowledge refers to a basic form of intelligence that we believe underlies our intellectual and communicative ability to cope with the tacit dimension of knowledge. The primordial knowledge model core group aims to build a computational model to understand primordial knowledge and incorporate the insights in artificial systems. We focus on the interactive knowledge process, attempting to develop a theory and technology for understanding and augmenting the mechanism of knowledge creation, transfer and application in a situated and embodied fashion through computer-mediated human-human communication.
Towards this end, we integrate the following approaches.
- Developing a technology for computational auditory scene analysis that allows robots to hear like humans.
- Developing a technology for capturing human behavior using computer vision.
- Developing a technology for creating, transferring, and applying situated knowledge content by building communicative artifacts that can participate in multimodal interactions.
- Developing a theory that can account for the biological mechanism of sensation.
- Investigation of computational properties of creation and transfer of primordial knowledge.
1. Computational auditory scene analysis computing for human-like robot audition
For symbiosis among people and robots/systems, one of critical issues is human-like robot/machine audition that can hear and listen to by its own ears. Since we hear not only voiced speech, musical sounds and environment sounds but also their mixtures in real-world environments, Robot audition requires computational auditory scene analysis (CASA) computing. Three fundamental functions for CASA are sound source localization, separation and recognition. In addition, robot audition should be independent on supposed acoustic environments and need less prior training. For the past two years, we have developed the open source software robot audition called "HARK" (old English standing for listen) and confirmed its feasibility by implementing it on several humanoid robots; a robot that can accept three simultaneous meal orders, and a robot referee for the Rock-Scissor-Paper Game by sounds. The basic technology for "Shoto-Taishi" (Prince Shotoku) robots has been established. We have also developed a music robot that can listen to music by its own ears, step and dance according to the beats it recognizes in real-time. We will apply CASA computing to spoken dialog systems, developmental communication that transforms signals to symbols and ground them into contexts, language acquisition by sound babbling like infants, and field informatics.
2. 3D Human Action / Behavior / Bearing / Facial Expression from 3D Video for Advanced Human Communication Interfaces
It is well known that non-verbal expressions such as (intentional) actions, (task-oriented) behavior, (subconscious) bearing, and facial expressions play a crucial role in human communication. In our increasingly advanced information society systems for information retrieval, analysis, expression and presentation should be human-centered, i.e. support a wide variety of people including those without any special knowledge or skills. We believe non-verbal human interfaces are an essential function of such human-centered information systems. With our current 3D video system we can measure 3D human body motion in a spatial resolution of a few millimeters at 30 frames per second. This study analyzes spatial and dynamic characteristics of 3D human actions, behavior, bearing and facial expressions from 3D video to help us to develop advanced human communication interfaces.
3. Situated Conversational Knowledge Management for Primordial Knowledge Model
The conceptual framework of situated knowledge management consists of a situated knowledge quantization for describing situated knowledge, a situated knowledge management environment that capture and reproduce situated knowledge quanta in the community, and situated knowledge agent that can interact with other agents in a situated fashion.
Preliminary results are four-fold. First, we investigated the method for allowing a conversational robot to effectively communicate intentions with people using nonverbal behaviors. Early results include a social robot that uses motion feedback to express its internal state and intention, and the use of interactive perception to establish and maintain joint intention. Second, we studied mutual adaptation -- a phenomenon we believe to exist between multiple learning agents being adapting with each other -- by taking a three stage approach consisting of a human-human WOZ experiment, a human-robot WOZ experiment, and a human-adaptive robot experiment. Instead of directly diving into the third stage, they observed in detail how people adapt with each other and how people increasingly build the protocols in communication. Third, we developed a generic, component-based platform for embodied conversational agents, i.e., interactive synthetic characters that have CG-based embodiment. GECA Protocol (GECAP), an XML based high-level communication protocol, is introduced to specify the protocol among the components. A script language (GSML) in introduced to specify an ECA’s behaviors. Fourth, we developed a system that can associate conversation quanta on varying places in the environment. An augmented reality system was developed to automatically setup the spatial coordinates in the real world from corresponding two-dimensional image point.
4. Format of information used by living organisms
External world is physical world, where objects, electromagnetic waves, air waves and temperatures exist. In contrast, we feel that objects have colors, sounds, hotness or coldness, smell and taste. These properties must be virtual reality produced in the brain. However, mechanism to correlate physical world and virtual reality has not been clarified. The purpose of the present study is to clarify the mechanism based on material. First, we review outline of the sensory system. Molecular cloning of temperature receptor shows that sensory receptor is a keyboard that make impulses as trigger signal when adequate stimuli is applied and that sensory system works like a keyboard instrument. When physical stimuli play the instrument, virtual reality may be produced in the brain. Second, we state strategy to clarify material basis of the virtual reality by analyzing unicellular organisms, such as Paramecium caudatum and Euplotes aediculatus.
5. Mathematical Knowledge and Computational Learning
Inductive inference is to derive general rules from concrete data, and has been called Computational Learning in the field of theoretical computer science. It also can be interpreted as transfer of information from teachers to learners. Inductive inference is generally understood as inference in the direction opposite to that of deductive inference, and it is believed in general that Mathematics is the most typical application of deductive inference. Against the belief, it has recently discovered that the proof of Hilbert’s basis theorem can be interpreted as computational learning. This means that In this research we investigate the process of generating and transferring mathematical knowledge, by clarifying the relation between transfinite ordinals and Computational Learning, that of topology and Computational Learning, and that between computational real functions and Computational Learning, and so on.
