Multimodal data available to us through the Internet and other electronic
media are explosively increasing both in number and in variety. To
handle such massive data for various purposes, new technologies are
in need of development. With this in mind, we have started investigating
a new methodology that allows us to discover from multimodal data
the information relevant to the purpose at hand (which is referred
to as “invariants”). To achieve this goal, we will study
several qualitatively different problems from different research areas,
in which multimodal data play a central role (e.g., visual/audio/text
processing, cognitive science, auditory perception and robotics) .
The problems are to be tackled with some of the recently developed
inductive learning machines loaded with an automatic model selection
mechanism (e.g., Penalized Logistic Regression Machines and Support
Vector Machines). The results will be analyzed in order to establish
a new methodology for discovery of invariants, which will be applicable
to problems across different areas of study .
