中央研究院 資訊科學研究所
Current Research Results
"ADF: an Anomaly Detection Framework for Large-scale PM2.5 Sensing Systems," IEEE Internet of Things Journal, To Appear.
Authors: Ling-Jyh Chen, Yao-Hua Ho, Hsin-Hung Hsieh, Shih-Ting Huang, Hu-Cheng Lee, and Sachit Mahajan

As the population density continues to grow in the urban settings, air quality is degrading and becoming a serious issue. Air pollution, especially fine particulate matter (PM2.5), has raised a series of concerns for public health. As a result, a number of large-scale, low cost PM2.5 monitoring systems have been deployed in several international smart city projects. One of the major challenges for such environmental sensing systems is ensuring the data quality. In this paper, we propose an Anomaly Detection Framework (ADF) for large-scale, real-world environmental sensing systems. The framework is comprised of four modules: 1) Time-Sliced Anomaly Detection (TSAD), which detects Spatial, Temporal, and Spatio-temporal anomalies in the real-time sensor measurement data stream; 2) Real-time Emission Detection (RED), which detects potential regional emission sources; 3) Device Ranking (DR), which provides a ranking for each sensing device; and 4) Malfunction Detection (MD), which identifies malfunctioning devices. Using real world measurement data from the AirBox project, we demonstrate that the proposed framework can effectively identify outliers in the raw measurement data as well as infer anomalous events that are perceivable by the general public and government authorities. Because of its simple design, ADF is highly extensible to other advanced applications, and it can be exploited to support various large-scale environmental sensing systems.
"Identifying Protein-protein Interactions in Biomedical Literature using Recurrent Neural Networks with Long Short-Term Memory," The 8th International Joint Conference on Natural Language Processing (IJCNLP 2017), November 2017.
Authors: Yu-Lun Hsieh, Yung-Chun Chang, Nai-Wen Chang and Wen-Lian Hsu

Accurate identification of protein-protein interaction (PPI) helps biomedical researchers to quickly capture crucial information in literatures. This work proposes a recurrent neural network (RNN) model to identify PPIs. Experiments on two largest public benchmark datasets, AIMed and BioInfer, demonstrate that RNN outperforms state-of-the-art methods with relative improvements of 10% and 18%, respectively. Cross-corpus evaluation also indicates that RNN is robust even when trained on data from different domains. These results suggest that RNN effectively captures semantic relationships among proteins without any feature engineering.
Current Research Results
Authors: Ming-Chin Chuang, Meng Chang Chen

Meng ChangChenAbstract:
In the vehicular network, the smart car provides variously network services via user equipment (UE). However, the services might be interrupted since the handover procedure of the UE is performed frequently. The existing handover decision schemes in cellular networks are mostly based on received signal strength (RSS) to make the handover decision. However, these schemes suffer from a ping-pong effect and, as a result, perform unnecessary handovers incurring extra signaling overhead and packet loss, which is even more exaggerated in the ultra-dense environment. The failure probability of the handover procedure will increase when the velocity of the car is fast. Therefore, in this paper we propose a navigation-assisted seamless handover (NASH) scheme for reducing unnecessary handovers and the ping-pong effect. NASH uses the massive Multiple Input Multiple Output (MIMO) technique to perform the bicasting scheme and then it utilizes coordination multi-point (CoMP) and carrier aggregation (CA) mechanisms in the handover procedure to avoid the packet loss problem and enhance throughput. Moreover, a dynamic time to trigger (TTT) mechanism is proposed according to the velocity of the car. The simulation results show that the proposed scheme has a better performance in terms of handover latency, packet loss, network throughput, handover failure probability, and the number of unnecessary handovers.
"Accurate audio-to-score alignment for expressive violin recordings," International Society of Music Information Retrieval Conference (ISMIR), October 2017.
Authors: Jia-Ling Syue, Li Su, Yi-Ju Lin, Yen-Kuang Lu, Yu-Lin Wang and Alvin W. Y. Su

An audio-to-score alignment system adaptive to various playing styles and techniques, and also with high accuracy for onset/offset annotation is the key step toward advanced research on automatic music expression analysis. Technical barriers include the processing of overlapped notes, repeated note sequences, and silence. Most of these characteristics vary with expressions. In this paper, the audio-to-score alignment problem of expressive violin performance is addressed. We propose a two-stage alignment system composed of the dynamic time warping (DTW) algorithm, simulation of overlapped sustain notes, background noise model, silence detection, and refinement process, to better capture the onset. More importantly, we utilize the non-negative matrix factorization (NMF) method for synthesis of the reference signal in order to deal with highly diverse timbre in real-world performance. A dataset of annotated expressive violin recordings in which each piece is played with various expressive musical terms is used. The optimal choice of basic parameters considered in conventional alignment systems, such as features, distance functions in DTW, synthesis methods for the reference signal, and energy ratios, is analyzed. Different settings on different expressions are compared and discussed. Results show that the proposed methods notably improve the conventional DTW-based alignment method.
Current Research Results
"Building/environment Data/information Enabled Location Specificity and Indoor Positioning," IEEE Internet of Things Journal, To Appear.
Authors: C. C. Li, J. Su, E. T. H. Chu and J. W. S. Liu

Jane Win ShihLiuAbstract:

The building/environment data and information system (BeDIS) described here is a part of infrastructure needed to support location-specific, active emergency preparedness and responses within large buildings. BeDIPS (Building/environment Data and information based Indoor Positioning System) is one of its components. BeDIPS can provide people in large buildings with sufficiently accurate location data. It is scalable, disaster resilient and is easy to configure, deploy and maintain. BeDIPS works without Internet and serves both smart phones and most legacy Bluetooth devices. The other component of BeDIS is BeDi mist, a virtual repository of data and information on the building, interior layouts and facilities. The mist uses micro data servers and smart gateways to deliver fine-scale, location-specific decision support data on a timely basis to hundreds and thousands of active devices and mobile applications.

Current Research Results
"A Position-Aware Language Modeling Framework for Extractive Broadcast News Speech Summarization," ACM Transactions on Asian and Low-Resource Language Information Processing, August 2017.
Authors: Shih-Hung Liu, Kuan-Yu Chen, Yu-Lun Hsieh, Berlin Chen, Hsin-Min Wang, Hsu-Chun Yen, and Wen-Lian Hsu

Extractive summarization, a process that automatically picks exemplary sentences from a text (or spoken) document with the goal of concisely conveying key information therein, has seen a surge of attention from scholars and practitioners recently. Using a language modeling (LM) approach for sentence selection has been proven effective for performing unsupervised extractive summarization. However, one of the major difficulties facing the LM approach is to model sentences and estimate their parameters more accurately for each text (or spoken) document. We extend this line of research and make the following contributions in this work. First, we propose a position-aware language modeling framework using various granularities of position-specific information to better estimate the sentence models involved in the summarization process. Second, we explore disparate ways to integrate the positional cues into relevance models through a pseudo-relevance feedback procedure. Third, we extensively evaluate various models originated from our proposed framework and several well-established unsupervised methods. Empirical evaluation conducted on a broadcast news summarization task further demonstrates performance merits of the proposed summarization methods.
Current Research Results
Authors: Ling-Jyh Chen, Yao-Hua Ho, Hu-Cheng Lee, Hsuan-Cho Wu, Hao-Min Liu, Hsin-Hung Hsieh, Yu-Te Huang, and Shih-Chun Candice Lung

As the population in cities continues to increase rapidly, air pollution become a serious issue from public health to social economy.  Among all pollutants, fine particulate matters (PM2.5) directly related to various serious health concerns, e.g., lung cancer, premature death, asthma, cardiovascular and respiratory diseases. To enhance the quality of urban living, sensors are deployed to create smart cities. In this paper, we present a participatory urban sensing framework for PM2.5 monitoring with more than 2,500 devices deployed in Taiwan and 29 other countries.  It is one of the largest deployment project for PM2.5 monitor in the world as we know until May 2017. The key feature of the framework is its open system architecture, which is based on the principles of open hardware, open source software, and open data. To facilitate the deployment of the framework, we investigate the accuracy issue of low-cost particle sensors with a comprehensive set of comparison evaluations to identify the most reliable sensor. By working closely with government authorities, industry partners, and maker communities, we can construct an effective eco-system for participatory urban sensing of PM2.5 particles. Based on our deployment achievements to date, we provide a number of data services to improve environmental awareness, trigger on-demand responses, and assist future government policymaking. The proposed framework is highly scalable and sustainable with the potential to facilitate the Internet of Things, smart cities and citizen science in the future.
Current Research Results
"Mixture of Gaussian Blur Kernel Representation for Blind Image Restoration," IEEE Transactions on Computational Imaging, To Appear.
Authors: Chia-Chen Lee , Wen-Liang Hwang

Blind image restoration is a non-convex problem involving the restoration of images using unknown blur kernels. The success of the restoration process depends on three factors: 1) the amount of prior information concerning the image and blur kernel; 2) the algorithm used to perform restoration; and 3) the initial guesses made by the algorithm. Prior information of an image can often be used to restore the sharpness of edges. By contrast, there is no consensus concerning the use of prior information in the restoration of images from blur kernels, due to the complex nature of image blurring processes. In this paper, we model a blur kernel as a linear combination of basic 2-D patterns. To illustrate this process, we constructed a dictionary comprising atoms of Gaussian functions derived from the Kronecker product of 1-D Gaussian sequences. Our results show that the proposed method is more robust than other state-of-the-art methods in a noisy environment, due to its increased signal-to-noise ratio (ISNR). This approach also proved more stable than the other methods, due to the steady increase in ISNR as the number of iterations is increased.
"Automatic Music Video Generation Based on Simultaneous Soundtrack Recommendation and Video Editing," ACM Multimedia Conference 2017, October 2017.
Authors: Jen-Chun Lin, Wen-Li Wei, James Yang, Hsin-Min Wang, and Hong-Yuan Mark Liao

An automated process that can suggest a soundtrack to a user-generated video (UGV) and make the UGV a music-compliant professional-like video is challenging but desirable. To this end, this paper presents an automatic music video (MV) generation system that conducts soundtrack recommendation and video editing simultaneously. Given a long UGV, it is first divided into a sequence of fixed-length short (e.g., 2 seconds) segments, and then a multi-task deep neural network (MDNN) is applied to predict the pseudo acoustic (music) features (or called the pseudo song) from the visual (video) features of each video segment. In this way, the distance between any pair of video and music segments of same length can be computed in the music feature space. Second, the sequence of pseudo acoustic (music) features of the UGV and the sequence of the acoustic (music) features of each music track in the music collection are temporarily aligned by the dynamic time warping (DTW) algorithm with a pseudo-song-based deep similarity matching (PDSM) metric. Third, for each music track, the video editing module selects and concatenates the segments of the UGV based on the target and concatenation costs given by a pseudo-song-based deep concatenation cost (PDCC) metric according to the DTW-aligned result to generate a music-compliant professional-like video. Finally, all the generated MVs are ranked, and the best MV is recommended to the user. The MDNN for pseudo song prediction and the PDSM and PDCC metrics are trained by an annotated official music video (OMV) corpus. The results of objective and subjective experiments demonstrate that the proposed system performs well and can generate appealing MVs with better viewing and listening experiences.
"Exploiting Asymmetric SIMD Register Configurations in ARM-to-x86 Dynamic Binary Translation," The 26th International Conference on Parallel Architectures and Compilation Techniques (PACT), September 2017.
Authors: Yu-Ping Liu, Ding-Yong Hong, Jan-Jan Wu, Sheng-Yu Fu, Wei-Chung Hsu

Processor manufacturers have adopted SIMD for decades because of its superior performance and power efficiency. The configurations of SIMD registers (i.e., the number and width) have evolved and diverged rapidly through various ISA extensions on different architectures. However, migrating legacy or proprietary applications optimized for one guest ISA to another host ISA that has fewer but longer SIMD registers through binary translation raises the issues of asymmetric SIMD register configurations. To date, these issues have been overlooked. As a result, only a small fraction of the potential performance gain is realized due to underutilization of the host’s SIMD parallelism and register capacity. In this paper, we present a novel dynamic binary translation technique called spill-aware SLP (saSLP), which combines short ARMv8 NEON instructions and registers in guest binary loops to fully utilize the x86 AVX host’s parallelism as well as minimize register spilling. Our experiment results show that saSLP improves the performance by 1.6X (2.3X) across a number of benchmarks, and reduces spilling by 97% (99%) for ARMv8 NEON to x86 AVX2 (AVX-512) translation.
"Virtual Persistent Cache: Remedy the Long Latency Behavior of Host-Aware Shingled Magnetic Recording Drives," ACM/IEEE International Conference on Computer-Aided Design (ICCAD), November 2017.
Authors: Ming-Chang Yang, Yuan-Hao Chang, Fenggang Wu, Tei-Wei Kuo, and David Hung-Chang Du

This paper presents a Virtual Persistent Cache design to remedy the long latency behavior of the Host-Aware Shingled Magnetic Recording (HA-SMR) drive. Our design keeps the cost-effective model of the existing HA-SMR drives, but at the same time asks the great help from the host system for adaptively providing some computing and management resources to improve the drive performance when needed. The technical contribution is to trick the HA-SMR drives by smartly reshaping the access patterns to HA-SMR drives, so as to avoid the occurrences of long latencies in most cases and thus to ultimately improve the drive performance and responsiveness. We conduct experiments on real Seagate 8TB HA-SMR drives to demonstrate the advantages of \\textit{Virtual Persistent Cache} over the real workloads from Microsoft Research Cambridge. The results show that the proposed design can avoid at least 94.93% of long latencies and improve the drive performance by at least 58.11%, under the evaluated workloads.
Current Research Results
"Three TF Co-expression Modules Regulate Pressure-Overload Cardiac Hypertrophy in Male Mice," Scientific Reports, August 2017.
Authors: Yao-Ming Chang, Li Ling, Ya-Ting Chang, Yu-Wang Chang, Wen-Hsiung Li, Arthur Chun-Chieh Shih*, and Chien-Chang Chen*

Arthur Chun-ChiehShihAbstract:
Pathological cardiac hypertrophy, a dynamic remodeling process, is a major risk factor for heart failure. Although a number of key regulators and related genes have been identified, how the transcription factors (TFs) dynamically regulate the associated genes and control the morphological and electrophysiological changes during the hypertrophic process are still largely unknown. In this study, we obtained the time-course transcriptomes at five time points in four weeks from male murine hearts subjected to transverse aorta banding surgery. From a series of computational analyses, we identified three major co-expression modules of TF genes that may regulate the gene expression changes during the development of cardiac hypertrophy in mice. After pressure overload, the TF genes in Module 1 were up-regulated before the occurrence of significant morphological changes and one week later were down-regulated gradually, while those in Modules 2 and 3 took over the regulation as the heart size increased. Our analyses revealed that the TF genes up-regulated at the early stages likely initiated the cascading regulation and most of the well-known cardiac miRNAs were up-regulated at later stages for suppression. In addition, the constructed time-dependent regulatory network reveals some TFs including Egr2 as new candidate key regulators of cardiovascularassociated (CV) genes.
Current Research Results
"On Space Utilization Enhancement of File Systems for Embedded Storage Systems," ACM Transactions on Embedded Computing Systems (TECS), April 2017.
Authors: Tseng-Yi Chen, Yuan-Hao Chang, Shuo-Han Chen, Nien-I Hsu, Hsin-Wen Wei, and Wei-Kuan Shih

In the past decade, mobile/embedded computing systems conventionally have limited computing power, RAM space, and storage capacity due to the consideration of their cost, energy consumption, and physical size. Recently, some of these systems, such as mobile phone and embedded consumer electronics, have more powerful computing capability, so they manage their data in small flash storage devices  (e.g., eMMC and SD cards) with a simple file system. However, the existing file systems usually have low space utilization for managing small files and the tail data of large files. In this work, we thus propose a dynamic tail packing scheme to enhance the space utilization of file systems over flash storage devices in embedded computing systems by dynamically aggregating/packing the tail data of (small) files together. To evaluate the benefits and overheads of the proposed scheme, we theoretically formulate analysis equations for obtaining the best settings in the dynamic tail packing scheme. Additionally, the proposed scheme was implemented in the file system of Linux operating systems to evaluate its capability. The results demonstrate that the proposed scheme could significantly improve the space utilization of existing file systems.
Current Research Results
"Supervised Learning of Semantics-Preserving Hash via Deep Convolutional Neural Networks," IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.
Authors: Huei-Fang Yang, Kevin Lin, and Chu-Song Chen

This paper presents a simple yet effective supervised deep hash approach that constructs binary hash codes from labeled data for large-scale image search. We assume that the semantic labels are governed by several latent attributes with each attribute on or off, and classification relies on these attributes. Based on this assumption, our approach, dubbed supervised semantics-preserving deep hashing (SSDH), constructs hash functions as a latent layer in a deep network and the binary codes are learned by minimizing an objective function defined over classification error and other desirable hash codes properties. With this design, SSDH has a nice characteristic that classification and retrieval are unified in a single learning model. Moreover, SSDH performs joint learning of image representations, hash codes, and classification in a point-wised manner, and thus is scalable to large-scale datasets. SSDH is simple and can be realized by a slight enhancement of an existing deep architecture for classification; yet it is effective and outperforms other hashing approaches on several benchmarks and large datasets. Compared with state-of-the-art approaches, SSDH achieves higher retrieval accuracy, while the classification performance is not sacrificed.
"Recognizing Offensive Tactics in Broadcast Basketball Videos via Key Player Detection," IEEE International Conference on Image Processing, September 2017.
Authors: T.Y. Tsai, Y. Y. Lin, H.Y. Mark Liao, and S. K. Jeng

We address offensive tactic recognition in broadcast basket- ball videos. As a crucial component towards basketball video content understanding, tactic recognition is quite challenging because it involves multiple independent players, each of which has respective spatial and temporal variations. Motivated by the observation that most intra-class variations are caused by non-key players, we present an approach that integrates key player detection into tactic recognition. To save the annotation cost, our approach can work on training data with only video-level tactic annotation, instead of key players labeling. Specifically, this task is formulated as an MIL (multiple instance learning) problem where a video is treated as a bag with its instances corresponding to subsets of the five players. We also propose a representation to encode the spatio-temporal interaction among multiple players. It turns out that our approach not only effectively recognizes the tac- tics but also precisely detects the key players.
"Distributed Compressive Sensing: Performance Analysis with Diverse Signal Ensembles," European Signal Processing Conference (EUSIPCO), August 2017.
Authors: Sung-Hsien Hsieh, Wei-Jie, Liang, Chun-Shien Lu, and Soo-Chang Pei

Distributed compressive sensing is a framework considering jointly sparsity within signal ensembles along with multiple measurement vectors (MMVs). The current theoretical bound of performance for MMVs, however, is derived to be the same with that for single MV (SMV) because the characteristics of signal ensembles are ignored. In this work, we introduce a new factor called “Euclidean distances between signals” for the performance analysis of a deterministic signal model under MMVs framework. We show that, by taking the size of signal ensembles into consideration, MMVs indeed exhibit better performance than SMV. Although our concept can be broadly applied to CS algorithms with MMVs, the case study conducted on a well-known greedy solver called simultaneous orthogonal matching pursuit (SOMP) will be explored in this paper. We show that the performance of SOMP, when incorporated with our concept by modifying the steps of support detection and signal estimations, will be improved remarkably, especially when the Euclidean distances between signals are short. The performance of modified SOMP is verified to meet our theoretical prediction.