# Deep Learning & Machine Learning

**Interpretable Machine Learning** Deep learning has been applied in a variety of areas including finance, medicine, image and speech recognition. For artificial neural networks, reliable factor analysis can increase confidence in prediction results, especially in the medical, financial, aviation, and defense sectors where a user requires confidence in results. We perform a study for deep learning approaches focusing on interpretability.

**Reinforcement Learning** Usually, a deep learning model predicts a certain value, such as probability or portfolio return. However, reinforce learning does not correspond to learning a value; instead, it corresponds to learning a policy. The reinforce learning focuses on how to obtain maximum rewards given an environment. We are applying such reinforce learning techniques for finance domain in particular.

**Generative models** Supervised machine learning models are used for predictions in many applications. By contrast, instead of predicting labels for given samples, generative models learn a data distribution to produce synthetic samples which are very similar to real ones. We are studying on novel generative models, in particular variants of GAN, for conditional generation of synthetic samples.

# Applications of Machine Learning

**Weather Forecast **Weather forecast has been conventionally conducted only by meteorological methods which bring difficulties to analyze effects of complex factors. In contrast, deep learning models have an advantage to make such problems solvable by considering inherent dependencies among variables. Therefore, the deep learning can show a possibility to improve the performance for weather forecast; in addition, we can easily extend it to a further research, such as weather-dependent energy demand forecasting.

**Computational Electromagnetics **We intend to innovate the design of nanophotonic material devices using novel AI technology by learning intuition and experience required for the design. By employing the finite-difference frequency-domain (FDFD) method, we built a database for the learning of AI models. With the database and the AI models, we can predict the characteristics of a certain nanophotonic device and automatically design a nanophotonic device with the desired characteristics.

# Financial Data Analysis

**Portfolio Optimization using Machine Learning **We use deep learning techniques to learn financial data and optimize portfolio for maximizing future returns. Our approach is conducted by the following processes: predict whether future prices go up or down, and find optimal parameters for a modern portfolio optimization theory. For such an approach, we use deep learning methodologies, such as LSTM, CNN, GAN etc.

**Ensemble Methods for Stock Market Prediction **Ensemble methods are commonly used to enhance the performance of prediction models. Such ensemble methods can also be used for the deep learning models, and they are a dominant method in many applications (e.g. natural language processing). Our objective is to develop better ensemble models using neural networks for prediction of stock prices.

# Bioinformatics

**Survival Analysis with Multiple Time Events **Conventional survival analysis has been widely studied and utilized to reveal relationships between an event and random variables or multiple events which can be randomly censored by hidden factors. Beyond the conventional analysis, we have been developing new methods, such as GAIT (Gene expression Analysis for Interval Time;* Bioinformatics* 2018).

**Synthetic Medical Data Generation **In order to develop a machine learning model for diagnosis, a large scale of database is required to train the model. However, due to the nature of medical data and privacy concern, it is hard to construct such database. Synthetic medical data generated by GAN models can be an alternative solution for the problem. We are studying on the generation of conditional time-series medical data, such as blood pressure data.

**Gene Network Estimation** Generally, networks can provide meaningful information between variables. For instance, gene networks can represent existence of relationships among genes, correlation coefficient of a connection, and hub genes which regulate many other genes. We are developing estimation methods to construct the gene networks.

# Applied Statistics

**Probability Estimation** Probability estimation is a fundamental task in statistical analysis. While numerous studies have been conducted for the problem, there are few methods for factor variables with many categories, such as genomic data and purchasable items in a mall, since they are newly emerging data from recent improvements on “Big data” technologies. We are studying on statistical methods for the probability estimation of such categorical variables.