
 Informationtheoretic learning theory/Statistical learning theory
 “What are the possibility and limitation of machine learning?” We take
informationtheoretic and statistical approaches to answer this
question. As for informationtheoretic learning theory, we study a
unifying methodology for model selection, representation learning,
change detection, highdimensional sparse learning, etc. on the basis
of the minimum description length principle. As for statistical
learning, we study new algorithm designs and theoretical analysis of
deep learning and kernel methods on the basis of statistical theory.
We also develop new optimization methods to run the machine learning
algorithms efficiently on the big data.
 Data Science Foundation
 We study methodologies for knowledge discovery from big data (anomaly
detection, network mining, embedding, etc.) Specifically we are
interested in discovering deep knowledge from latent spaces. We aim at
building a new field called “Symptomatics”, in which we detect signs
of latent changes in future from data streams.
 Data Science Applications
 We study effective data science methodologies by applying machine
learning and data mining technologies to real complex data. The
applications include economics, financial data analysis, medical data
analysis, marketing, SNS data analysis, failure detection, spatial
data mining, security, etc. We often collaborate with industrial
companies to solve real data science problems.
