Dimensionality reduction is a technology to transform data from a high-dimensional space into a low-dimensional space so that the low-dimensional representation retains some meaningful and latent properties of the original data. In our Lab, we focus on dimensionality reduction methods that preserve the local relationships in high-dimensional space, such as locality preserving projection (LPP) and Laplacian Eigenmap-based methods.
Clustering is an important technique for exploratory data analysis with the objective of grouping unlabeled data objects in the same cluster which are similar to each other. We are focusing on the algorithms that are closely related to the above dimensionality reduction methods such as spectral clustering. We are also developing clustering technologies in collaboration with the medical, life, and materials informatics fields.
Advanced deep learning technology
In deep learning, the performance depends on what kind of learning method is used. In our laboratory, instead of the usually used stochastic gradient descent method, we have developed a learning method that uses non-negative matrix factorization. We are also developing a hybrid method of the non-negative matrix factorization and the stochastic gradient descent. We are also working on a high-performance implementation for parallel computing environments.
Application to life science and medical data
We are developing various machine learning application technologies for life science and medical data in collaboration with researchers, medical doctors, companies, etc. We handle various data such as gene expression data, metabolite data, and genomic variation data, and are working on the practical application of the developed algorithms and theories.