| 山田 暉 | M, 1回目発表 | 自然言語処理学 | 松本 裕治, 中村 哲, 中田 尚, 新保 仁, 進藤 裕之 |
|
Title: Multiple Emotion Detection in Text Dialogues
abstract: In recent years, the research of emotion detection in dialogues has gained popularity. This task is to detect emotions of a present utterance using not only present utterance's features, but also previous utterances' features (and sometimes future utterances' features). Emotions we call in these tasks are, for example, joy, sadness, fear, anger, etc. Most of the related works until now were to detect only one emotion. However, 1 utterance might represent multiple emotions. Therefore, we propose the model of Autoencoder multiple-emotion detection in text dialogues. We use EMTC dataset which is annotated multiple emotions to conversational transcripts. This research can be useful for customer support in chat dialogues. Language of the presentation: Japanese | |||
| CHANG XIN | M, 1回目発表 | ロボティクス | 小笠原 司☆, 向川 康博, 金出 武雄（客員）, 高松 淳, 伍 洋 |
|
title: Towards Real-time Multi-Object Visual Tracking
abstract: With the rapid growth of deep-learning based artificial intelligence, multiple computer vision tasks, e.g. object detection and single-object tracking, are handled with better accuracy and in more unified approaches. However, multi-object tracking, which is a higher-level task, hasn't been dealt with well using deep neural networks. The common tracking-by-detection paradigm considers multi-object tracking as a two-stage problem - 1) detection and 2) association among detections. In this paradigm, the object detector has no knowledge about targets that are being tracked, and would get false negatives due to occlusion. Based on suboptimal detection results, the association algorithm can hardly find best matches. More importantly, GPU computation power couldn't be fully utilized with non-learning association algorithms, thus limiting the processing speed to lower than 5 fps. In this work, we focus on speeding up multi-object tracking algorithms using recurrent neural networks, making the multi-object tracking technology ready for real-life applications. By integrating detector and matcher into one framework, we discover approaches to object re-detection with given knowledge of targets. language of the presentation: English | |||
| DANG CHENYU | M, 1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, 金出 武雄（客員）, 荒川 豊, 伍 洋, 諏訪 博彦 |
|
title: Deep Multi-Cue For Taxi Demand Prediction
abstract: Taxi demand prediction is an important part of intelligent transportation systems in a smart city. How to build a more accurate prediction model to reduce resource wasting of empty taxi on streets is becoming a hot topic. Most of research used large-scale taxi demand data from taxi requesting services(e.g.,Japan Taxi,DiDi), but it depends on whether passenger use the application or not. Due to that, I proposed a novel way to find out potential passengers based on advanced computer vision technology(Object detection,Human Pose Estimation). We firstly detect if people around interesting postion then estimate their pose to analysis if they want to take taxi. This method can be universal and multi-purpose. However,for real-time application,computational speed is the key factor. In this work, I combined CenterNet detector and MobileNet V1 based human pose estimator and got a faster result. language of the presentation: English | |||
| PHETCHAI PONPAT | M, 1回目発表 | サイバーレジリエンス構成学 | 門林 雄基, 林 優一, 安本 慶一, 妙中 雄三 |
|
Title: Differentiating DDoS from Flash Crowd Traffic
Abstract: Distributed Denial of Service (DDoS) attack is one of the costless to learn and easy to execute attack, which have been existing for an extensive period of time. However, recently, there are many cases that proved the attack is still persistently successful. One of the famous examples is the Mirai IoT botnet. The DDoS attacks uses a simple concept of flooding packets to the target device or network until it reaches the physical or logical limits, which resulted in inaccessible of legitimate uses to the device or network. Therefore, Denial of Service (DoS). To be Distributed DoS, the key to success is the number of devices under control and the packets being send to destination without being blocked or detected by the target. Hence, a way to avoid that is to forge the source IP addresses. Comparing with a similar type of traffic called flash crowd, resulted from Slashdot effect, where the users intended to get the information at the same time. Flash crowd have no intention to forge the IP addresses. With the believes that the righteous user should have the right to access, where the DDoS should not. We are currently researching on a method to distinguish between the DDoS attack from flash crowd traffic by using the traffic behaviour with regards to the geolocation information. Since there are many approaches and effort on detecting DDoS attack, but rarely did focus on the geolocation distribution and unique characteristic of the 2 types of traffic. If we can distinguish the two types of traffic, we could not only minimise the error in DDoS detection for various other models, but to detect the DDoS itself as well. In this presentation, we will introduce you to some of the endless efforts that researchers in this field have been working on to counter against the DDoS attacks and provide the reason why much more attention is needed to prevent such kind of attacks. Together with this, we will discuss on our aspect and concept to the resolution of distinguishing DDoS and flash crowd by deriving geolocation information. Moreover, the expected goals and our current progress will also be reported here. Last but not least, we will conclude by linking back to the overview and explain how our approach contribute to solving the main problem. Language of the presentation: English | |||
| CHAUQUE EUCLIDES TELES TOMAS | M, 1回目発表 | 情報基盤システム学 | 藤川 和利, 安本 慶一, 新井 イスマイル |
|
title: Urban Traffic Prediction From Mobility Data Using Deep Learning
abstract: Accurate and timely traffic flow information is important for the successful deployment of intelligent transportation systems. Over the last few years, traffic data have been exploding, and we have truly entered the era of big data for transportation. Existing traffic flow prediction methods mainly use shallow traffic prediction models and are still unsatisfying for many real-world applications. In this talk, I am going talk about how deep learning techniques are being used by researches to improve models accuracy, focusing on a stacked autoencoder model. language of the presentation: English | |||
| 田中 雄基 | M, 1回目発表 | 生体医用画像 | 佐藤 嘉伸, 末次 志郎（バイオ）, 大竹 義人, スーフィーマーゼン, 日朝 祐太 |
|
title: Automated Approach for Assessment of Acetabular Coverage Rate of Diseased Hip Joint Based on CT Images by Using Convolutional Neural Network
Abstract: Acetabular dysplasia is an ageing-related disease, whose prevalence covers 80% among hip osteoarthritis cases in Japan. Acetabular dysplasia occurs when the acetabulum (hip socket) is shallow and does not provide sufficient coverage of the femoral head (ball), causing instability of the hip joint. Therefore, it is important to understand the acetabular coverage rate (ACR) for diagnosis, preoperative planning and post-operative evaluation of treatment outcomes. However, the acetabular coverage is usually performed on two-dimensional (2D) radiographs despite the effects of pelvic tilt that makes it difficult to compare between diseased and normal hip joints. The purpose of this study is to develop an automated approach for three-dimensional (3D) assessment of ACR of the femoral head based on CT images. In this presentation, I introduce an anatomical landmark detection approach using Convolutional Neural Network (CNN) with preliminary results on the estimation of the Center Edge Angle (CEA), which is clinically used as an indicator of ACR. language of the presentation: Japanese | |||
| TALUSAN JOSE PAOLO VICTORIA | D, 中間発表 | ユビキタスコンピューティングシステム | 安本 慶一, 中島 康彦, 荒川 豊, 諏訪 博彦, 水本 旭洋 |
|
Title: Distributed Machine Learning over Edge Computing Environments for Smart Connected Community Services
Abstract: Internet-of-Things (IoT) is leading the movement towards smart connected communities. Pervasive applications rely on the data generated by IoT devices to provide services over the cloud. However, the cloud struggles to meet service-level objectives (SLOs) such as availability and response time for more complex tasks. Edge and fog computing paradigms build upon this by moving processing nearer the data sources. This work presents distributed machine learning over edge nodes by using a middleware based on the Information Flow of Things framework. By offloading the processing to the edge, latency is reduced and availability is increased. While these edge environments are able to perform basic computation tasks. In order to realize more complex tasks such as machine learning training and inference, multiple edge devices need to be able to efficiently utilize computational resources. In this work, we show how distributed machine learning can be done over edge computing environments. We show how such a middleware can be used to train and recognize activities of daily living, proving that the tradeoffs favor processing gain over communication costs. language of the presentation: *** English *** | |||
| ARAYA KIBROM DESTA | M, 2回目発表 | 情報基盤システム学 | 藤川 和利, 林 優一, 新井 イスマイル |
|
title: Intrusion Detection in Controller Area Network Using Nueral Networks
abstract: Nowadays, vehicles are equipped with multiple Electronic Control Units(ECUs) each of which communicate with one another using a specification called Controller Area Network (CAN). CAN provides its own share of benefits in modernizing automobiles, but it also brought along a security issue to the automotiveindustry. CAN bus does not have any mechanism for encrypting or authenticatingCAN payloads. As a countermeasure against this drawbacks we have experimentedon identifying intrusions in the CAN bus using Long Short-Term Memory Networks(LSTM). LSTM networks are trained to predict a forthcoming payloads and relatedattributes by looking at information that has already appeared in the CAN bus atsome instant in time. The predicted values are compared with actual values, thatare either sent by an intruder or the benign ones, and depending on how close ourprediction is with the actual payload, we managed to effectively identify anomalies inan acceptable accuracy rate, up to 98%. We have tested our method with a varietyof attacks to the CAN bus, and demonstrated how effective our detection method is. language of the presentation: English | |||