コロキアムB発表

日時: 12月15日(水)3限(13:30~15:00)


会場: L3

司会: Chen Na
上原 誠 M, 1回目発表 ソーシャル・コンピューティング 荒牧 英治, 中村 哲, 若宮 翔子, 劉 康明
title: Visualization of Confirmation Bias by Hoax Mention Tweets in Social Media
abstract: Social media has now become an inevitable part of our lives, but the spread of misinformation such as fake news and hoaxes is a social problem. Confirmation bias is the tendency to collect information that is convenient to one's beliefs and to disregard conflicting information. However, it is difficult to reduce it, because the original belief is further strengthened when conflicting beliefs are provided. In this study, we aim to create a visualization tool for confirmation bias in social media to help users collect information objectively.
language of the presentation: Japanese
発表題目: ソーシャルメディアのデマ言及ツイートによる確証バイアスの可視化
発表概要: 近年,ソーシャルメディアの普及に伴い,フェイクニュースやデマといったいわゆる偽の情報の拡散が社会問題となっている.ソーシャルメディア上の情報拡散の要因の一つに,自分の信念にとって都合の良い情報を収集したり,対立する情報を軽視する確証バイアスがあるが,対立する信念を示されるとさらに元の信念が強化されるといった問題もあることから,その解消は難しい.本研究では,ユーザ自身が確証バイアスを認知することが客観的な情報判断をするうえで重要になると考え,ソーシャルメディア上に存在する確証バイアスを分析,可視化するツールの構築を目指す.
 
KIYOMOTO BARBARA M, 1回目発表 知能コミュニケーション 中村 哲, 渡辺 太郎, 須藤 克仁

title: Word Emotion Classification Based on Use Frequency in Movie Genres

abstract: In Sentiment Analysis, there are many datasets with data annotated with positive/negative sentiment. There are relatively less existing datasets with data annotated with more specific emotion labels. This research hopes to explore the idea of automatically generating an emotion lexicon by classifying word emotion based on word frequency across various movie genres. The distributional hypothesis posits that words with similar meaning will occur closely together in similar situations. Based on this hypothesis, the assumption is made that words with similar emotion will occur more frequently together in the same movie genres.

language of the presentation: English

 
RUMMAN MAHFUJUL ISLAM M, 1回目発表 計算システムズ生物学 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘

title: Feature extraction and unsupervised clustering of cellular phenotypes in histopathological images of pancreatic cancer using information maximization.

abstract: Pathological images are microscopic images of various types of tissues. Since the color of the images depends on staining methods, we can integrate different staining methods to obtain the features of various cell types. In this study, we will construct a model to embed these images into a latent space to extract biological features from those images using convolution. The latent features will be utilized for clustering based on information maximization. Overall, the reconstruction process and clustering process is jointly trained to minimize the total loss using a neural network.

language of the presentation: English  

 
YANG ZIWEI M, 1回目発表 計算システムズ生物学 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘

title:  An unsupervised learning framework for identifying cancer subtypes 

abstract: Cancer is very crucial and accurately classifying the subtypes of cancer is a critical step towards a more precise diagnosis and effective cancer treatment clinically. With merging the deep learning method, the results are promising. Especially, the deep generative model (DGM) learns the data attributes itself in different subtypes without using the luxury manual label. For instance, there are some works designed on top of variation autoencoder (VAE) frameworks. VAE utilizes a lower latent space to represent the data feature space. 

However, there is an issue in cancer data that the cancer data is a treasure and the dataset usually is very small. Therefore, using that small dataset is hard to fit a Gaussian strong assumption.In this work, we introduced VQ-VAE to escape the VAE issue and would like to construct an unsupervised learning framework for identifying cancer subtypes. 

language of the presentation: English 

 
高野 誠也 M, 1回目発表 情報セキュリティ工学 林 優一, 岡田 実, 安本 慶一, 藤本 大介, Youngwoo Kim
title: A Study on Evaluation Method of Speech Sound Focusing on Acquisition of Audio Information leakage via Electromagnetic Emission
abstract: Threats of audio information leakage by analyzing electromagnetic (EM) emission from input/output devices such as speakerphones have been reported. Previous studies evaluated audio information leakage via EM emission by measuring chirp sounds and their leakage spectrogram. However, actual speech sounds are different from the characteristics of chirp sounds; these evaluation methods may overestimate the target of human voices. This study proposes a method to evaluate information leakage using the probability of recovering words from a leaked electromagnetic emission as an index, using a human voice as the sound to be evaluated.
language of the presentation: Japanese
発表題目: 電磁波を介した音声情報漏えいの情報取得性に着目した発話音声の評価に関する研究
発表概要: スピーカーフォンなどの入出力機器から漏えいする電磁波を測定、解析することで音声情報を復元する攻撃が報告されている.これまで、音声が電磁波を介して漏えいするかの評価には、チャープ音などの規則的な音を取得し、スペクトログラムを用いた音信号のパターンを見る手法が用いられていた.しかし、実際の会話などの音声は不規則であり、規則的な音を用いた評価は過評価となる可能性がある.そこで本研究では、評価対象となる音声を人の声とし,漏えいした電磁波から情報を復元する確率を指標とした情報漏えい評価手法を提案する.