A5068376487- Medical Image Segmentation Techniques
- Image Retrieval and Classification Techniques
- Advanced Vision and Imaging
- Advanced Image and Video Retrieval Techniques
- Machine Learning and Data Classification
- Image Processing Techniques and Applications
- Advanced Image Processing Techniques
- Machine Learning and Algorithms
- Tensor decomposition and applications
- Blind Source Separation Techniques
- Sparse and Compressive Sensing Techniques
- Domain Adaptation and Few-Shot Learning
- Image and Signal Denoising Methods
- Face and Expression Recognition
- Phonocardiography and Auscultation Techniques
- Robotics and Sensor-Based Localization
- Advanced Numerical Analysis Techniques
- Computer Graphics and Visualization Techniques
- Soft Robotics and Applications
- Imbalanced Data Classification Techniques
- Digital Image Processing Techniques
- Neural Networks and Applications
- Advanced Clustering Algorithms Research
- AI in cancer detection
- Image and Object Detection Techniques
IBM Research - Tokyo
2025
Nagasaki University
2014-2024
NEC (Japan)
2020-2022
RIKEN
2016-2021
The University of Tokyo
2008-2020
RIKEN Center for Advanced Intelligence Project
2017-2018
National Institute of Technology, Kisarazu College
2015-2016
Tokyo Institute of Technology
2013-2014
Chiba University
2003-2011
Hiratsuka City Hospital
1994-1996
Overparameterized deep networks have the capacity to memorize training data with zero \emph{training error}. Even after memorization, loss} continues approach zero, making model overconfident and test performance degraded. Since existing regularizers do not directly aim avoid loss, it is hard tune their hyperparameters in order maintain a fixed/preset level of loss. We propose direct solution called \emph{flooding} that intentionally prevents further reduction loss when reaches reasonably...
Most of the semi-supervised classification methods developed so far use unlabeled data for regularization purposes under particular distributional assumptions such as cluster assumption. In contrast, recently from positive and (PU classification) risk evaluation, i.e., label information is directly extracted data. this paper, we extend PU to also incorporate negative propose a novel approach. We establish generalization error bounds our show that decrease with respect number without are...
Foreground object identification can be considered as anomaly detection in a redundant background. This paper proposes unsupervised deep learning of foreground objects on the basis prior knowledge about spatio-temporal sparseness and low-rankness background scenes. The proposed framework trains U-Net model to encode decode sparse batches input images with low-rank backgrounds, by minimizing combination nuclear ℓ1 norms loss function. approach is similar subtraction based robust principal...
Honeycombing on high-resolution computed tomography (HRCT) is a distinguishing feature of usual interstitial pneumonia and predictive poor outcome in lung diseases (ILDs). Although fine crackles are common ILD patients, the relationship between their acoustic features honeycombing HRCT has not been well characterized.Lung sounds were digitally recorded from 71 patients with findings chest HRCT. Lung analyzed by fast Fourier analysis using sound spectrometer (Easy-LSA; Fukuoka, Japan). The...
Squared-loss mutual information (SMI) is a robust measure of the statistical dependence between random variables. The sample-based SMI approximator called least-squares (LSMI) was demonstrated to be useful in performing various machine learning tasks such as dimension reduction, clustering, and causal inference. original LSMI approximates pointwise by using kernel model, which linear combination basis functions located on paired data samples. Although proved achieve optimal approximation...
The goal of binary classification is to identify whether an input sample belongs positive or negative classes. Usually, supervised learning applied obtain a rule, but in real-world applications, it conceivable that only and unlabeled data are accessible for learning, which called from (PU learning). Furthermore, practice, distributions likely differ between training testing due to, example, time variation domain shift. covariate shift dataset situation, where covariates (inputs) testing, the...
A five degree-of-freedom (DOF) miniature parallel robot has been developed to precisely and safely remove the thin internal limiting membrane in eye ground during vitreoretinal surgery. simulator determine design parameters of this robot. The robot's size is 85 mm × 100 240 mm, its weight 770 g. This incorporates an emergency instrument retraction function quickly from case sudden intraoperative complications such as bleeding. Experiments were conducted evaluate performance master-slave...
Abstract Background We aimed to develop an artificial intelligence (AI)-assisted oral cytology method, similar cervical cytology. focused on the detection of cell nuclei because ratio cytoplasm increases with increasing malignancy. As initial step in development AI-assisted cytology, we investigated two methods for automatic blue-stained cells cytopreparation images. Methods evaluated usefulness sliding window method (SWM) and mask region-based convolutional neural network (Mask-RCNN)...
In PU learning, a binary classifier is trained from positive (P) and unlabeled (U) data without negative (N) data. Although N missing, it sometimes outperforms PN learning (i.e., ordinary supervised learning). Hitherto, neither theoretical nor experimental analysis has been given to explain this phenomenon. paper, we theoretically compare (and NU) against based on the upper bounds estimation errors. We find simple conditions when NU are likely outperform prove that, in terms of bounds,...