Satoshi Iizuka

User-controllable Example-based Video and Image Contents Processing via Deep Neural Networks

Researcher

Satoshi Iizuka

Outline

I will establish a novel deep learning-based framework for user-controllable, example-based video and image contents processing. In content creation industries, most products are required to be of high quality and controllable by humans, but current deep neural network-based methods to achieve this have not yet been established. This project aims to develop a flexible and reliable framework by constructing not only a novel neural network-based image processing architecture that effectively leverages example data but an automatic example acquisition system.

Yohei Oseki

Building machines that process natural language like humans

Researcher

Yohei Oseki

Outline

In this project, I will build machines that process natural language like humans, by integrating the cognitive/brain sciences of language (a branch of natural science) and natural language processing (a branch of artificial intelligence). Specifically, language processing will be reverse-engineered through comparing human language processing measured via crowdsourcing and functional neuroimaging, on one hand, and machine language processing implemented as symbolic generative models and deep neural networks, on the other hand.

Kensuke Okada

Development of statistical methods for transparent operation of achievement tests

Researcher

Kensuke Okada

Outline

This research project sought to develop a framework of the item response model that allows the disclosure of test items after each administration of achievement tests. Statistical estimation methods that have desirable properties such as consistency are also developed for the models within the framework. As the test is more vulnerable to security threats due to the advancement of optical and telecommunication technologies, traditional test operation that keeps test items secret has reached a turning point. By developing a novel mathematical approach to assessments, this study builds a sustainable, transparent, and credible scheme of operation for achievement tests in the years to come.

Shunsuke Ono

A Mathematical Data Analysis Framework to Bridge the Gap between Measurements and Knowledge Discovery

Researcher

Shunsuke Ono

Outline

By mathematically incorporating various properties and conditions (sensing knowledge) associated with measurement objects, processes, and environments, we aim to develop fundamental technologies for overcoming the degradation and incompleteness inherent in measurement data and analyzing signal information of sufficient quality and quantity for knowledge discovery. The following three technical issues will be addressed. (1) mathematical assurance based on sensing knowledge, (2) highly reliable and efficient parameter selection, and (3) methodologies for reimporting and utilizing acquired knowledge for data analysis.

Yuya Sasaki

Creation of Explanable Technologies on Bias in Graph Data

Researcher

Yuya Sasaki

Outline

Graph data is widely used for many applications such as knowledge bases and social network services, but often includes discriminatory biases. Despite fairness in AI technologies becoming more important than mere accuracy, there are no studies on the formalization of biases in graph data and the discovery/deletion of such biases. In this project, I contribute to a fair world with AI technology through creating explainable technologies on biases in graph data. More concretely, I aim to explainably formalize biases in graph data and develop efficient methods for discovering biases and deleting discriminatory biases.

Matthew･James HOLLAND

Machine learning with guarantees under diverse risk measures

Researcher

Matthew･James HOLLAND

Outline

Modern machine learning methods are designed, often implicitly, with optimal average performance in mind; this is embodied in the ubiquity of the risk (expected loss) in the guarantees attached to algorithms in traditional statistical learning theory. While in many cases strong performance on average is desirable, there are countless other ways of evaluating (off-sample) performance, and committing to the expected value is, and should be recognized as, a substantial value judgement. In this work, we explore new notions of performance beyond the expected loss, seeking to obtain a new class of performance metrics which effectively balances flexibility and interpretability of the new metrics with the computational and theoretical tractability of algorithms designed to optimize the new criteria.

Takashi Matsubara

Geometric deep learning for ensuring desired properties by design

Researcher

Takashi Matsubara

Outline

For safety and reliability, artificial intelligence is expected to ensure properties of targeted systems, such as symmetry, structure, causality, stability, and laws of physics. Deep learning does not ensure these properties, whereas manually-designed mathematical models have a limitation in quantitative accuracy. In this project, I improve and generalize the concept of geometric deep learning, whose hypothesis space is designed by geometric viewpoints, thereby balancing the qualitative properties and quantitative accuracy.

Hitomi Yanaka

Logical Inference System between Data and Text For Decision Making

Researcher

Hitomi Yanaka

Outline

To support evidence-based decision making, it is desirable to build a framework for representing richer meanings of data and text and handling inference between them. The goal of this study is to combine logic and machine learning approaches to build a logical inference system between data and text to support human decision making. We consider how to obtain unified meaning representations for data and text and perform inference between them in a sophisticated manner. The study also aims to develop an inference system that outputs a proof process easy for humans to understand.

Daisuke Yokogawa

Transformative feature extraction and utilization based on the chemical knowledge

Researcher

Daisuke Yokogawa

Outline

Most of the machine learnings in chemistry focus on a single target, such as a reactivity prediction in organic chemical reactions and inorganic material design. Therefore, the application of the machine learning model trained for one target to another target is still difficult. In this study, I propose a machine leaning model that is applicable to multitask problems by focusing on the transformative features in chemistry.