This website was originally created as a portal site of a research project, “Discovering Deep Knowledge from Complex Data and Its Value Creation” (2011-2015) led by Professor K. Yamanishi (Univ of Tokyo), which was a subproject of a JST (Japan Science and Technology Agency) grant titled “Advanced Core Technologies for Big Data Integration” (JST CREST Grant Number JPMJCR1304).
One of the main research agendas of the project is best summarized by the concept of “Latent Dynamics.” Although the JST project itself has been completed, we are now planning to launch a new research project built upon the achievements of Latent Dynamics.
The world is complex and ever-changing. In the Big Data era, expectations have been growing that big data will provide us with the capability of analyzing and eventually predicting the society’s future.
One useful approach to analyzing complex dynamics is to introduce a hierarchy. Semiconductor physics, for example, analyzes the movement of billions of quantum-scale particles as the primary objective. Due to the interaction among elementary particles, finding the exact solution of the system’s dynamics is hopeless even in a tiny piece of semiconductor. In reality, however, semiconductor engineers can precisely control how electrons move collectively in a computer device, almost only with macroscopic theories of electric circuits, without being distracted by quantum mechanical equations. This is an example where hierarchical abstraction drastically simplifies the analysis.
Analyzing a complex system consisting of a huge number of evolving subsystems is one of the biggest challenges in data science. Due to the interaction among millions of individual subsystems, the principle of superposition does not work anymore. Also, unlike semiconductor physics, we generally do not have the precise law governing individuals’ behavior. We do not even know what kind of abstraction is possible and which level of the hierarchy is useful to our application of interest. However, we believe that there will be a scientific methodology to systematically extract useful knowledge from evolving complex systems, including not only social systems but also many physical and cyber-physical systems. The remarkable success of internet-age business applications such as search engines and online advertising campaigns seems to support this view.
Latent Dynamics is the term that best summarizes our research agenda. Hidden deep behind the apparent complex behaviors, a certain layer of abstraction may exist and it may dramatically advance our understanding of the system. A relatively simple representation in the latent layer may help capture the major driving force of the whole dynamics. Hidden Markov models and linear dynamical systems are classical tools for describing latent dynamics. Recurrent neural networks (RNNs) are another useful tool that has attracted renewed attention recently as a nonlinear extension of the classical models. However, aside from a few specific areas such as speech recognition and natural language processing, we are still only halfway to the ultimate goal.
Our main research questions include:
- How to learn a latent representation from observable data in an automated manner.
- How to detect changes over time in the latent representation that may trigger a disruptive change of the system.
- How to control the outcome of the system through a latent representation.
Along with advancing scientific research, we will actively engage with real-world business applications of latent dynamics.
The original Japanese version was written in 2011. This English version was added by T. Ide on 07/04/2020 for the purpose of archiving past activities.
IBSML (Information-Based Induction Sciences and Machine Learning; The leading academic society in machine learning in Japan)