Takashi Ishizuka

A study of advancing computational complexity theory for search problems to quantum computation

Researcher

Takashi Ishizuka

Outline

The purpose of this study is to unravel the difference in computational power between quantum and classical computation from the aspect of search problems. Computational complexity theory for search problems is a research field that has influenced a wide realm such as economics and machine learning. I will attempt to establish a new studying method for quantum computation by extending the knowledge of computational complexity theory to quantum computation theory, following the knowledge of the theory of computational complexity for search problems that have been studied on classical computation.

Kaito Ito

Integrating control and optimal transport theories for control of large-scale systems

Researcher

Kaito Ito

Outline

With the development of IoT technologies, the problem of controlling large-scale systems composed of many agents, such as swarm robots and transportation systems, has become more and more important. This project focuses on developing methods to efficiently control and transport an agent population to a desired distribution shape. The main approach is to integrate and develop ideas of optimal transport and control theories. In addition, by explicitly considering the uncertainty of the data used for control, this project aims to make the methods applicable to real-world problems.

Koji Inoue

Human-targeted evaluation based on multimodal behaviors in spoken dialogue

Researcher

Koji Inoue

Outline

A model that indirectly evaluates the “human-likeness” of users interacting with a dialogue system is realized. The metric is the similarity between the multimodal behaviors used with a spoken dialogue system and those used in real human-human dialogue, which is called human-targeted evaluation. A second model is also proposed that explains the reasoning for this similarity result. Finally, the proposed evaluation scheme is implemented in a spoken dialogue system, and a dialogue experiment is conducted to investigate its validity.

Satoshi Kura

Verification of machine learning algorithms via category theory and automated verification

Researcher

Satoshi Kura

Outline

With the recent development of machine learning, various demands have emerged for machine learning algorithms, such as privacy, fairness, and safety. I aim to obtain new methods for automated verification of machine learning algorithms written as probabilistic or differentiable programs by (1) generalizing existing automated verification methods such as program logic and refinement type systems using categorical semantics, and (2) concretizing them in a setting tailored to probabilistic or differentiable programs.

Kazuhiro Kurita

Development of fundamental technologies for maximal subset enumeration with ordering constraints

Researcher

Kazuhiro Kurita

Outline

In this research, I aim to solve the gap between theory and practice of enumeration algorithms. This gap comes from the increase of computational cost due to the number of outputs. Enumeration requires time that depends on the number of outputs. Thus, enumeration requires a huge computational cost even if we use theoretically efficient enumeration algorithms, like constant amortized time enumeration algorithms. In order to make the number of outputs adjustable, we focus on enumeration with ordering constraints. If we can solve enumeration problems with ordering constraints, then it achieves a trade-off between completeness of output and computational cost. For this reason, I study the ordering constrained enumeration problems as fundamental technologies for useful enumeration.

Kazuki Koga

Improving shape data analysis based on computational harmonic analysis

Researcher

Kazuki Koga

Outline

A precise method for evaluating similarities between three-dimensional shape data is of increasing importance not only because of traditional areas such as CAD and physics simulations but because of recent remarkable progress in data science. In this project, by regarding geometric objects as Schwartz distributions, I develop a novel framework of three-dimensional shape data analysis from low-level building blocks using tools in computational harmonic analysis, and aim to establish a new standard of the research area with both accuracy and efficiency.

Yuka Kotorii

Research on string-shaped matter by using topology

Researcher

Yuka Kotorii

Outline

The purpose of this research is to elucidate the structure of complicated string-shaped matter and relationship between the structure and the matter. Therefore, my goal is to provide a topological indicator that can be applied to various string-shaped matter. Furthermore, I carry out basic research on topology by using topological invariants and random topology, and also construct bases for applied research.

Ken Kobayashi

Development of scalable algorithms for mixed-integer semidefinite optimization

Researcher

Ken Kobayashi

Outline

This project aims to develop modeling techniques and algorithms for solving mixed-integer semidefinite optimization problems. First, I devise a modeling technique to reformulate the problem into a problem minimizing a convex function over an integer lattice. Then, I design a cutting-plane algorithm to solve the reformulated problem, and improve its computational efficiency by exploiting the sparsity included accelerate the algorithm by exploiting the sparsity of the problem.

Shohei Satake

RIP matrices breaking the square-root bottleneck and related combinatorics

Researcher

Shohei Satake

Outline

In this research, we first aim to construct new explicit RIP matrices breaking the square-root bottleneck, which is an important and challenging topic in the theory of compressed sensing. Next, based on the RIP of the constructed matrices, we aim to establish a unified approach to combinatorial areas such as pseudo-randomness, Ramsey theory and additive combinatorics.

Kota Shiba

Efficient architecture of 3D-stacked AI chips

Researcher

Kota Shiba

Outline

A highly efficient AI inference architecture based on a 3D-stacked SRAM is studied in this project. The main challenge is to realize an architecture that can efficiently process pruned sparse neural networks by actively utilizing the low-latency and random-access properties of the 3D-SRAM. I aim to create a next-generation computing system that can scale even in the post-Moore era by coordinating integrated circuit technology and AI algorithms, i.e., 3D system integration and AI model pruning.

Ryoma Shinya

A measure theoretic approach to formal language theory

Researcher

Ryoma Shinya

Outline

Regular measurability is a natural notion of “asymptotic approximation by regular languages”, which can be naturally characterized in terms of a measure theory called Caratheodory condition. In this research, I first clarify the structure of “C-measurable regular languages” for a local variety C of regular languages, and then extend its theory to more stronger language classes such as “regular-measurable context-free languages”. I also discuss the applicability of the notion of regular measurability to other fields.

Youyou Cong

Developing a music generation tool based on type theory

Researcher

Youyou Cong

Outline

In the programming languages community, it is well-known that types correspond to propositions, and programs correspond to proofs. In this work, I develop a music generation tool based on this correspondence. The idea is to represent music composition rules as types, and use a program synthesizer to generate well-typed music. This makes it possible to automatically generate music whose theoretical correctness is mathematically proved.

Yusuke Tanaka

Bayesian learning of neural network surrogates for simulating physical systems

Researcher

Yusuke Tanaka

Outline

A goal of this study is to establish a Bayesian framework for learning neural network surrogates of physical systems from noisy and sparse data. In Hamiltonian mechanics, the system’s dynamics is represented as orbits in phase space. In this study, I newly introduce a probabilistic generative model of orbits in phase space and derive an algorithm for learning the surrogates on the basis of Bayesian procedures. The effectiveness of the proposed model is demonstrated using synthetic data that are generated from known physical systems.

Taira Tsuchiya

Development and analysis of computationally-efficient and high-performance policies for a wide range of sequential decision-making problems

Researcher

Taira Tsuchiya

Outline

A sequential decision-making problem is a mathematical model in which a user makes a certain kind of decision based on successively observed information. In recent years, a large number of decision choices are available, and the variety of available information is increasing. However, it is known that existing policies with theoretical guarantees are inherently difficult to be executed in practical settings and have poor numerical performance. In this research, I aim to develop an algorithm that solves these problems.

Tomoki Nakamaru

Jupyter Notebook extension for efficient trial and error

Researcher

Tomoki Nakamaru

Outline

Trial-and-error is a common strategy to find the best hyper-parameter values and data-preprocessing methods in the development of machine-learning applications. This project aims at developing an automatic cache technique to streamline such a trial-and-error process. The developed technique will be implemented as a Jupyter Notebook extension that can be used in practice.

Tomohiko Nakamura

Sampling-frequency-independent deep learning framework for audio media processing

Researcher

Tomohiko Nakamura

Outline

Audio source separation is a technique of separating individual sources from a mixture audio, and it is often used for preprocessing of audio applications. To build a source separation model that can be used as a versatile preprocessor, various acoustic conditions (for example, sampling frequency) required by possible downstream tasks should be handled. Although conventional source separation models based on deep neural networks work well only at a trained sampling frequency, they are difficult to work with sounds of untrained sampling frequencies. In this study, interpreting deep neural networks from a signal processing viewpoint, I develop layers independent of sampling frequency to establish a more versatile deep learning framework for audio media processing.

Ryuichiro Hataya

Towards human-complementary generative models using energy-based models

Researcher

Ryuichiro Hataya

Outline

Deep generative models can generate data that nearly imperceptible from real data. On the other hand, human can leverage knowledge of rare events. This research project aims to integrate human knowledge to deep generative models, especially energy-based generative models, to enable more diverse and flexible generation.

Natsuki Hamanishi

Research on low-cost spatio-temporal interaction technology for observing human motion that bridges the gap of experience between video and face-to-face

Researcher

Natsuki Hamanishi

Outline

The use of videos to learn physical movements such as choreography is a well-known method. The ability to rewind 10 seconds and compare multiple videos is an advantage of video learning environments. In order to adapt these features to face-to-face learning environments, I develop software technology that enables the motion information of face-to-face person editable. While respecting both video and face-to-face learning modalities, I attempt to construct an information environment realizing interaction with motion information, spatio-temporal data in real space.

Yosuke Higuchi

End-to-end speech recognition based on the hierarchy of linguistic units

Researcher

Yosuke Higuchi

Outline

End-to-end automatic speech recognition (E2E-ASR) is a deep learning framework for realizing direct speech-to-text conversion. This project aims to improve the recognition performance of E2E-ASR, focusing on the hierarchy of linguistic units included in the text information. By training an E2E-ASR model such that it gradually processes linguistic information (e.g., “character” to “word”), I expect the model to permit extracting more sophisticated linguistic features from the input speech. Our model will help develop spoken language understanding, such as in conversational dialogue systems, where a system needs not only to recognize but also to understand what the users said.

Masahiro Fujisawa

Developing machine learning with robustness and generalization

Researcher

Masahiro Fujisawa

Outline

“Robustness” and “Generalization” are essential for reliable machine learning applications in the real world by reducing the risk of prediction errors. However, it is a known fact that there is a trade-off between them, and the theoretical analysis and development of mitigation methods are conducted only on the specific models, tasks, and contaminated data. In this project, I will try to reveal universal theoretical mechanisms for the trade-off between robustness and generalization and establish a novel machine-learning algorithm that can simultaneously acquire robustness and generalization performance.

Kenta Hoshino

Reachability analysis for systems on space of probability measures and applications to deep learning

Researcher

Kenta Hoshino

Outline

This project aims to develop a framework of reachability analysis for stochastic dynamical systems. The reachability analysis is known as the analysis for guaranteeing the performance of systems. The project aims to extend the reachability analysis for guaranteeing the performance of stochastic systems on the space of probability measures. Additionally, the analysis for stochastic systems will be adapted to the analysis of deep learning.

Hikaru Hoshino

Multiscale modeling for electricity tariff design

Researcher

Hikaru Hoshino

Outline

This project aims to develop a modeling method for profitability analysis of Distributed Energy Resources (DERs) on the demand side for the purpose of the appropriate design of electricity tariffs. Specifically, by developing a method for dealing with the trade-off between the short-term decision of the customer on the operation of DERs and the long-term decision on the installation of DERs, I provide a novel tool to intuitively perform a sensitivity analysis of the profitability of DERs.

Seito Matsubara

Human-machine dynamics based on a motion induction system model

Researcher

Seito Matsubara

Outline

When human movement is assisted or controlled with a muscle actuator, such as electrical muscle stimulation, a critical issue is the integration of such induced movement with the person’s motion intention and how this movement then affects their motor control. Towards achieving optimal integration and reducing feelings of artificiality and enforcement, I will develop a mathematical model of human perception of synthesized movements by voluntary and induced one through psychophysical experiments. And also, I will develop an adaptive electrode selection device that can stimulate the most suitable area on the muscle for electrical muscle stimulation when users are moving. Finally, I propose a mathematical model and design method for a system that makes motion intervention transparent (motion induction system).

Akihiro Mizutani

Realization of information theoretically secure quantum information processing with practical devices

Researcher

Akihiro Mizutani

Outline

Unlike classical cryptography that is currently used, quantum cryptography achieves the ultimately secure communication and is expected to be put into practice. Quantum computing, which is another central topic in quantum information science, has computational power that surpasses that of classical computing, and recently its cloud usage has been started. In this project, I will conduct the research that realize more efficient quantum cryptographic communication and that enable the secure usage of quantum computers in the cloud. Through the research, I aim to realize efficient and secure quantum information processing using practical devices that can be developed with current technology.

Maria Katarina Larsson

Computer-assisted wood craft

Researcher

Maria Katarina Larsson

Outline

Solid wood is homogenous and anisotropic; its material properties vary with the grain pattern. I propose to approximate this volumetric grain structure based on images of the visible exterior surfaces. Then I will create a material-aware interface that leverages the grain information to optimize structural performance, adapt the fabrication process, and preview the appearance of the to-be-fabricated object. I hope that this material-aware interface will be used to create higher quality wooden artifacts via smarter material use, without consuming more material resources.

Tianxiang Liu

Developing efficient algorithms for structured constrained optimization problems with applications to machine learning

Researcher

Tianxiang Liu

Outline

The research on constrained optimization problems in which constraints are represented by nonconvex nonsmooth functions is difficult and has just begun. In this project, I aim to build structured constrained optimization models for a wide class of problems including many important application problems in fields such as machine learning and system identification, and then develop efficient algorithms for the models.