| 後上 正樹 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 池田 和司, | 荒川 豊（客員教授）, | 松田 裕貴 | ||
|
title:
Abstract: [ Private due to patent pending ] language of the presentation: Japanese | ||||||||
| 福田 修之 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 池田 和司, | 荒川 豊（客員教授）, | 松田 裕貴 | ||
|
title: Statistical analysis between sleep status and occupational health questionnaire toward detection of depression symptoms 発表題目: 健常者のうつ兆候検知に向けた睡眠状況と労働衛生指標の統計分析 abstract: In a previous study, 60 office workers working at 5 general companies equipped wearable devices, which are widely used and could be easily introduced by companies, for 2~3 weeks. Also, we collected daily sleep data and questionnaire responses to measure worker's mental status. Based on these data, our findings reveals that two groups of high and low scores of DAMS(Depression and Anxiety Mood Scale) questionnaire, which are depression, affirmation, and anxiety mood, at waking-up time clarified the different sleep structure and appearance. 発表概要: 近年，企業に勤める労働者の行動変容を目的として，日々の行動や心理・生理状態を認識する研究が盛んに行われている． 本研究の目標は，企業に勤めるオフィスワーカの日々の心身状態を把握し，それぞれにあったサポートすることで行動変容を促し，快適性向上や病気の予防・早期発見による健康管理を行うことである． 先行研究では，一般企業5社に勤めるオフィスワーカ60人に対して，一般的に普及しており企業や個人が容易に導入可能なウェアラブルデバイスを2~$\sim$~3週間の装着し，日々の睡眠データと労働者の精神状態を測定するアンケート回答を収集した． language of the presentation: Japanese | ||||||||
| KHONG THI THU THAO | D, 中間発表 | コンピューティング・アーキテクチャ | 中島 康彦, | 池田 和司, | 中田 尚, | Tran Thi Hong, | 張 任遠 | |
|
title: Bayesian Convolutional Neural Network and its application againstadversarial attacks
abstract: Bayesian Convolutional Neural Network (BCNN) is a combination of a Bayesian Neural Network (BNN) and a Convolutional Neural Network (CNN). In BCNN, the weight is a probabilistic distribution, which is single point estimate in CNN. The stochastic components of BNN helps the model obtain the prediction as a distribution and evaluate the uncertainty of the prediction. Due to the weight uncertainty, BCNN becomes a good defense against adversarial attacks that are gradient based attacks. There are many defense methods, e.g. randomness, ensemble, and adversarial training that can improve the robustness of CNNs with respect to adversaries. Among them, adversarial training is the most outstanding defense. However, the training cost is a big challenge to adversarial learning. In this research, we propose a new defense algorithm called Bayes without Bayesian Learning, which does not add the training phase in resisting to adversarial attacks. Our method is applied to pretrained CNN models on both natural and perturbed data, and improves the robustness of them under strong attacks. For example, on ImageNet with Resnet50 network, we achieve 57% accuracy in the top5 under PGD attack and 3% improvement when combining our approach with adversarial training. language of the presentation: English | ||||||||
| 川崎 明宙 | M, 2回目発表 | 生体医用画像 | 佐藤 嘉伸, | 池田 和司, | 末次 志郎, | 大竹 義人, | Soufi Mazen, | 上村 圭亮 |
| title: *** Statistical Analysis of Sacral Bone Density using a CNN-based Atlas Creation *** abstract: *** Fragility fracture of the sacrum has been an issue for elderly people. Research has been limited for the sacrum partly because its shape is complex with large inter-subject variation. Also, large-scale statistical analysis of its shape and density distribution has been limited mainly due to the computational load in establishment of the voxel correspondences, i.e., deformable registration. In this study, we employed a convolutional neural network (CNN)-based deformable registration algorithm in the analysis of the sacral bone. The algorithm we employed, VoxelMorph (Dalca et al. Med Image Anal 2019), is characterized as an unsupervised algorithm where no ground truth deformation field is required. The algorithm also allows to create a conditional deformable template, which is a volume image exhibiting smallest deformation field from all samples with specific attributes (e.g., age, sex, etc.), which in short represents the “average” (or “centroid”) image among the specific age and sex group. We applied it on a database consisting of 837 CTs (149 males, 688 females, 58.14 ± 14.73 y.o.) of the pelvis region, where the sacral bone was segmented and masked. We computed the templates corresponding to ages of 20 to 80 for male and female. The templates visually illustrated reduction of the bone density with aging in both male and female. The quantitative analysis showed that the average CT value over the sacrum region was reduced from 135.73 HU to 39.88 HU for 20 y.o. template to 80 y.o. template.*** language of the presentation: *** Japanese*** | ||||||||