Research
Research Topics and Relevant Publications. Please see this list of publications.
Our current research focuses on the following topics: i) Converting human experience, intuition, and techniques into language using data, ii) Developing sustainable operation techniques for artificial intelligence systems, and iii) Fundamental techniques for media information processing, pattern recognition, and machine learning that serve as the foundation for the aforementioned topics.
Sustainable Construction and Operation of AI Systems for Human Decision Making Support
Our research is focused on developing methodologies for sustainable artificial intelligence (AI) systems that can support human decision-making. Specifically, we aim to enable the early operation of AI systems, even in the absence of big data, and to adaptively grow these systems to perform well in individual environments, particularly in the context of infrastructure maintenance and decision support for primary industry workers in fields such as livestock and fisheries. While deep learning technology has advanced significantly in the era of big data, applying end-to-end deep learning methods to real-world problems without big data is not always practical. Hence, we aim to address the question, ``What can we do when big data is not available in the Big Data era?''
Our research focuses on the following approaches:
Data-driven Feature Extraction: We use deep neural networks constructed with relevant large-scale data to extract essential features for a target task. Key techniques we employ include feature representation learning, self-supervised learning, and transfer learning.
Data Assimilation: We incorporate expert knowledge into the model. Key techniques include state-space models, time series analysis, and Bayesian modeling.
Human-in-the-Loop Machine Learning: We adapt the system to the target environment by making unlabeled data available through the use of human power. Key techniques we employ include semi-supervised learning, active learning, and crowdsourcing.
As a decision support system, it is essential that the AI system is intuitive for the user and that the results can be easily explained. Thus, we emphasize the importance of integrating inductive and deductive approaches to effectively utilize domain knowledge and data.
Smart Maintenance
We are conducting research and development on technologies that can detect signs of potential failure in industrial machinery in advance to achieve cost-effective maintenance.
(Click here for details of the project. )
Precision Fisheries
We are conducting research and development on technologies for acquiring, analyzing, and predicting information directly related to fishing decisions and fisheries resource management (e.g., technologies for predicting daily catches and identifying productive fishing grounds).
(Click here for details of the project. )
Precision Livestock Farming
We are conducting research and development on technologies to help livestock farmers make better decisions. Our focus is on monitoring the condition of livestock using video information, which includes identifying signs like calving and estrus.
(Click here for details of the project. )
Medical and Nursing Support
We are conducting research and development on AI technology that can aid in communication with children who have profound intellectual and multiple disabilities, with a focus on predicting the emotional state and intentions of the child using video images, along with a rationale for the prediction.
Crowdsourcing Utilization
We are developing a framework called Tutti to facilitate the utilization of crowdsourcing, with the goal of enabling crowdsourcing to be used in a manner similar to programming.
(Click here for details of the project. )
Speech and Acoustic Signal Processing
We are developing methodologies to create speech processing systems that can deliver robustly high performance even with unknown inputs to the system.
Speech Recognition
Speech recognition is a process of estimating the spoken words from an audio signal. We are currently focused on developing and refining this basic technology, with a particular emphasis on approaches that are robust to environmental variations.
(Click here for details of the project. )
Speech Enhancement
Speech enhancement is a process of separating and extracting clear speech sounds from noisy or reverberant environments. We are currently focused on developing and refining this basic technology and applying it to robot auditions.
(Click here for details of the project. )
Speaker Recognition
Speaker recognition is a process of identifying ``who spoke (and when)'' from speech utterances. We are currently focused on developing and refining this basic technology, with a particular emphasis on approaches that are robust to environmental variations.
(Click here for details of the project. )
Speech Synthesis
Speech synthesis technology involves generating speech signals either from text or by converting existing speech signals. We are currently exploring the optimal methods for evaluating sound quality in the era of crowdsourcing.
Singing Voice Recognition
We are advancing basic technologies to enable the scoring of singing voices beyond just pitch or technique as well as those to evaluate the similarity of singing voices for song recommendation.
Image and Video Information Processing
We are conducting research and development on fundamental technologies for image and video processing applications.
Object Identification
We are developing and refining basic techniques for identifying objects in images and video, with a particular focus on approaches that are robust to environmental variations and noise in labels.
(Click here for details of the project. )
Object Detection and Tracking
We are developing and refining highly efficient and accurate technologies for detecting specific objects in images and video, as well as tracking moving objects in video, with a focus on user interaction.
(Click here for details of the project. )
Image Recognition and Security
We are developing and refining the basic technology to enable the analysis of images and videos (e.g., making machine learning available) while maintaining privacy protection.
(Click here for details of the project. )
Age and Gender Recognition
Gender and age estimation is a technique used to estimate a person's gender and age based on facial images. We are developing and refining this basic method, with a particular emphasis on techniques for acquiring features that align with human intuition.
(Click here for details of the project. )