About the Laboratory
Our laboratory addresses a wide range of challenges grounded in machine learning and statistical learning theory.
We foster an environment where undergraduate and graduate students can engage actively in research, and we pursue collaborative projects with researchers both inside and outside the university.
Machine Learning Methodology
Most modern AI technologies, including deep learning, are built on the field of machine learning. Classical AI relied on manually crafted rules and hand-tuned systems. Machine learning, by contrast, takes a data-driven approach, inferring models statistically from data. Since the 2000s, it has been demonstrated across many domains that data-driven models can sometimes far surpass manually constructed systems, attracting enormous societal attention. Our laboratory conducts research on machine learning for diverse tasks from two perspectives: foundational mathematical research forming a universal basis, and practical data analysis applications.
Optimization
Black-box Optimization via Bayesian Models
We study techniques that use data-driven predictive models for optimization. Many real-world problems can be formulated as optimization problems — for example, finding the best material composition to maximize battery performance in next-generation energy storage, or efficiently searching for compounds with promising pharmacological properties in drug discovery. In such settings, acquiring data itself can be costly, making it critical to search efficiently with limited observations. We investigate Bayesian probabilistic model-based optimization from both foundational mathematical and applied perspectives.
Automated Machine Learning (AutoML)
While machine learning is widely known to be applicable to many problems, practical deployment often requires expert-level tuning. Manually determining which model and configuration to use for a given task and dataset is extremely costly. The field of AutoML seeks to automate this process of finding appropriate machine learning settings. Beyond simply maximizing predictive accuracy, there is increasing societal demand for optimization that reflects diverse human preferences.
Interpretable AI
Interpretable Machine Learning Models
Discovering which factors are important within machine learning predictive models is crucial in many contexts. For example, in genomic data analysis, one might need to identify the most influential genes or gene interactions — out of tens of thousands — related to a disease or biological phenomenon of interest. We study models with explicitly interpretable internal mechanisms as well as methodologies for interpreting the outputs of AI systems.
Knowledge Discovery from Structured Data
When data has complex structure that is difficult to represent as simple numerical values — such as molecular graphs — the problem of identifying important substructures becomes both harder and more important. In such cases, combinatorial explosion makes exhaustive enumeration on modern computers practically infeasible. Our laboratory researches machine learning models that can efficiently deliver highly interpretable results even for complex structured data.
Scientific Applications
Materials Informatics
Developing new materials with desired properties requires significant cost. The endeavor to efficiently discover new materials using data science is known as materials informatics. Our laboratory addresses various machine learning-based materials analyses in collaboration with materials scientists.
Bioinformatics
Data-driven analysis plays diverse roles in modern biology. For example, large-scale genomic data have become increasingly accessible, and statistical analysis of genomic data to understand biological phenomena is now commonplace. In collaboration with biologists, our laboratory researches frameworks for accelerating biological discovery through machine learning — including protein data analysis and related problems.