How Intelligent Can Computers Be?
Theory, Algorithm, and Application of Machine Learning
With the dramatic performance improvement of information and communication technology,
intelligent information processing that can be done only by humans
is becoming possible also by computers.
Under the theme of "how intelligent can computers be?",
Sugiyama Laboratory is working on various research topics
related to intelligent data analysis, called machine learning,
in the field of artificial intelligence.
- Construction of Learning Theory
Generalization is the ability to be able to cope with unknown situations,
and is indispensable for computers to behave intelligently.
We are theoretically investigating the mechanism of acquiring the generalization capability
based mainly on probability and statistics.
- Development of Learning Algorithms
Machine learning involves various subjects such as
supervised learning (learning from input-output paired data),
unsupervised learning (learning from input-only data),
and reinforcement learning (learning through interaction with an environment).
We are developing practical and theoretically motivated machine learning algorithms.
- Application of Machine Learning Technologies to Real-World
Growth and spread of the Internet and sensor technologies
allow us to collect a huge amount of data in engineering and fundamental sciences
such as documents, audio, images, movies, e-commerce, electric power, medicine, and biology.
We are collaborating with industry partners and applying
state-of-the-art machine learning technologies to solving real-world challenging problems.
Computational Imaging and Vision from Space
Yokoya Laboratory is working on image processing for computational imaging and image analysis based on computer vision and machine learning. In particular, we work on intelligent information processing to automatically extract map information, such as land cover labels and elevation models, from remote sensing images acquired by spaceborne and airborne sensors.
- Computational Imaging
Computational imaging, which integrates sensing and computation, allows us to acquire information that cannot be obtained by hardware alone and to overcome hardware limitations, such as resolution and noise. Based on machine learning, optimization, and signal processing, we build mathematical models and develop algorithms to recover unknown original signals from incomplete observation data.
- Remote Sensing Image Analysis
Remote sensing enables us to observe places that are inaccessible to humans; however, it is difficult to collect enough training data due to the limitations of field surveys and visual interpretation. We work on mapping and 3D reconstruction by using synthetic data from simulations and inaccurate labels with low collection costs as training data. We also work on data fusion based on deep learning to handle multimodal data obtained from different spaceborne sensors in an integrated manner.
- Towards a Sustainable Future
We promote projects to solve global issues, including environmental problems, climate change, large-scale natural disasters, and food problems. Our goal is to contribute globally to the realization of the SDGs by solving real-world problems, such as assessing building damage during disasters, estimating biomass and carbon stocks in forests, and mapping crop types, in collaboration with related institutions and researchers in Japan and overseas.
Towards Practical and Reliable Machine Learning
Ishida Laboratory started in 2021 and is currently conducting fundamental research to develop machine learning algorithms. For example, we are building algorithms to learn from weak supervision and regularizers that can alleviate overfitting. We aim to make machine learning more practical and reliable through our research.
- Weakly Supervised Learning
To perform ordinary supervised learning successfully, we need a large amount of labeled data. However, it is often costly to collect appropriate supervision or difficult to collect due to business constraints. To utilize an alternative source of inexpensive supervision, we are working on weakly supervised learning such as complementary-label learning and positive-confidence learning.
- Overfitting and Regularization
In real-world applications, we often encounter many challenges that hinder us from achieving high prediction performance, such as learning from a small amount of data or learning with label noise. To cope with these situations, we are working on making machine learning more robust and proposing regularization methods to alleviate overfitting.
- Evaluation of Machine Learning Models and Data
Given a trained machine learning model, how will we know if there is any room left for improvement? If the model has reached the best achievable error, it is meaningless to continue aiming for error improvement. Hence, knowing the best achievable error is helpful since it enables us to compare it with the trained model’s error. This can also be used as a criterion of the dataset difficulty. We are developing methods that can be used to evaluate models and datasets, such as directly estimating the best achievable error.