The diffusion model, a new generative modeling paradigm, has achieved significant success in generating images, audio, video, and text. It has been adapted for sequence-to-sequence text generation (Seq2Seq) through DiffuSeq, termed the S2S-Diffusion model. Existing S2S-Diffusion models predominantly rely on fixed or hand-crafted rules to schedule noise during the diffusion and denoising processes. However, these models are limited by non-contextualized noise, which fails to fully consider the characteristics of Seq2Seq tasks. In this paper, we propose the Meta-Diffu$B$ framework—a novel scheduler-exploiter S2S-Diffusion paradigm designed to overcome the limitations of existing S2S-Diffusion models. We employ Meta-Exploration to train an additional scheduler model dedicated to scheduling contextualized noise for each sentence. Our exploiter model, an S2S-Diffusion model, leverages the noise scheduled by our scheduler model for updating and generation. Meta-Diffu$B$ achieves state-of-the-art performance compared to previous S2S-Diffusion models and fine-tuned pre-trained language models (PLMs) across four Seq2Seq benchmark datasets. We further investigate and visualize the impact of Meta-Diffu$B$'s noise scheduling on the generation of sentences with varying difficulties. Additionally, our scheduler model can function as a "plug-and-play" model to enhance DiffuSeq without the need for fine-tuning during the inference stage.

Generative Adversarial Networks (GANs) have been popularly researched in natural language generation, so-called Language GANs. Existing works adopt reinforcement learning (RL) based methods such as policy gradients for training Language GANs. The previous research of Language GANs usually focuses on stabilizing policy gradients or applying robust architectures (such as the large-scale pre-trained GPT-2) to achieve better performance. However, the quality and diversity of sampling are not guaranteed simultaneously. In this article, we propose a novel meta-learning-based generative adversarial network, Meta Exploration GAN (MetaEx-GAN), for ensuring the quality and diversity of sampling (sampling efficiency). In the proposed MetaEx-GAN, we develop an explorer trained by Meta Exploration to sample from the generated data to achieve better sampling efficiency. MetaEx-GAN employs MetaEx first applied to Language GANs to achieve better performance. We also propose a critical training method for MetaEx-GAN on the NLG task. According to our experimental results, MetaEx-GAN achieves state-of-the-art performance compared with existing Language GANs methods. Our experiments also demonstrate the generality of MetaEx-GAN with different architectures (involving GPT-2) and how MetaEx-GAN operates to improve Language GANs.

TeraSort [7] is a standard MapReduce sort which is applied as a benchmark to measure the time to sort terabytes of randomly distributed data. TeraSuffix [5] adopts TeraSort to construct suffix array for NGS (Next-Generation Sequencing). When using TeraSuffix to construct suffix array for NGS, as the intermediate data of the MapReduce framework contains multiple copies of suffixes, the shuffle of the intermediate data between map and reduce become a bottleneck. For a suffix can be represented by its index to the NGS Reads data, it's no need to record suffix as the intermediate data and thus reduce the shuffle time. Disk-based Indexed TeraSuffix [6] adopts this index structure to represent a suffix and stores the NGS Reads data on disk. However, when constructing the suffix array, reduce tasks still need lots of random access from disk to retrieve suffixes for further processing. The massive disk I/O operations become a bottleneck. To increase the efficiency of the Disk-based Indexed TeraSuffix, in-memory lookup cloud (MLC) is proposed in this paper to store the NGS Reads data on the memory of remote servers in a cloud. When a map/reduce task needs to retrieve a suffix, it can access the suffix from the memory of MLC through the network. Experimental tests were performed to show that the access of suffix through network is outperformed than from disk. Experiments were also performed on Amazon Elastic MapReduce with the sequence of 20Gbp-Grouper (about 20Gbytes). It showed that the proposed architecture reduces the pre-processing time of data replication and the processing time of reduce tasks by 58% and 8%, respectively. It improves both the time and the space efficiency of TeraSort for constructing suffix array.

This paper proposes a new investment strategy with deep-learning market prediction for mutual fund portfolio optimization. Our strategy uses the capital asset pricing model (CAPM) that applies macroeconomic factors to predict whether the market is bull or bear. Then, we develop a robo-advisor (RA) to predict future market, optimize portfolio and automate investment. Experiments use 22 years' data of S&P500 and mutual funds of U.S. to validate our strategy. Results show that the accuracy of our market prediction method can reach 84.3% and the rate-of-return of our RA is 13.S7%. Our model is more accurate and profitable than other algorithms.

According to the latest investigation, there are 1.7 million active social network users in Taiwan. Previous researches indicated social network posts have a great impact on users, and mostly, the negative impact is from the rising demands of social support, which further lead to heavier social overload. In this study, we propose social overloaded posts detection model (SODM) by deploying the latest text mining and deep learning techniques to detect the social overloaded posts and, then with the developed social overload prevention system (SOS), the social overload posts and non-social overload ones are rearranged with different sorting methods to prevent readers from excessive demands of social support or social overload. The empirical results show that our SOS helps readers to alleviate social overload when reading via social media.

Framing is a phenomenon that is studied and debated widely in sociology and political science. It refers to the manner in which audiences interpret information and justify their claims or activities. The subconscious influence of framing might lead to opinion changes and social movements. However, multi-frame classification on microblogging data has not yet been investigated. In this study, we aim to classify a large number of posts into frames. We describe in detail the implementation of a new algorithm for multi-frame classification tasks called Frame Dispatcher, which aims to classify microblogging data into frames. In our experiments, we extracted over 15,000 posts from approximately 200 Facebook fan pages concerning an anti-curriculum student movement. The experimental results show that Frame Dispatcher can classify microblogging data into frames efficiently and effectively.

Since a complete Chinese font has typically several thousand or more Chinese characters and symbols, and most of them are much more complicated than English alphabets, it takes a lot of time and efforts for even professional font engineers to create a Chinese font. Although several attempts had been made to synthesize Chinese characters from strokes and components, it is still not easy to synthesize so many Chinese characters at one time. In this paper, we present an easy and fast solution for an ordinary user to create a Chinese font of his or her handwriting style. We adopt the approach: to synthesize Chinese characters using components extracted from the user's handwritings. In the preprocessing phase, we built a Web interface for crowds to label the positions and sizes of components of every Chinese character in the target character set. The standard Kai font was selected as a reference. We also devised an algorithm to find a small subset of Chinese characters having all required components to synthesize other Chinese characters. To create a personal handwriting font, with commonly-used 3,914 traditional Chinese characters, a user only has to handwrite 400 or so Chinese characters on a pad. One character by one character, our system can track every stroke, recognize and extract components from the user's handwritings. Then, every target Chinese character is synthesized from the extracted components, by placing them properly according to their position and size information. The experiment results show that although manually fine-tune is still required for few synthesized Chinese characters, users can create a Chinese font of their personal handwriting styles more easily and quickly.

The purpose of this paper is to develop a method to recommend products to customer via mobile devices. Collaborative recommendation is known as an effective way to recommend products. In this paper, we use the concept of collaborative recommendation to develop Mobile-Oriented Catalog (MOC). The proposed method is made from aggregating similar purchasing records to optimize combination of goods on mobile devices. This paper illustrates how to design attractive and collaborative catalog to recommend items by using Latent Dirichlet Allocation (LDA) based self-adaptive genetic algorithm (LDA-SAGA). LDA-SAGA is consisted of topic modeling concept and self-adaptive genetic algorithm. We use LDA as our topic modeling algorithm to construct MOC as a result that it is the simplest topic model. Our experimental evaluation on synthetic and real data shows that using preference as topic concept is effective. LDA-SAGA is especially outstanding with large number of customers and products. Finally, we compare the MOC which is used on mobile application (APP) of Amazon with the one used on Taobao and discuss the characteristics of their design. Different design of user interface on APP can lead to different scope of fitness value which is capable of explaining different market strategies of Taobao and Amazon.

Efficient use of energy is an important research topic of the smart grid. Load monitoring is an integral part of energy management, convenient information, communication technology, and sensor applications. So far, many monitoring techniques have been developed, and non-intrusive load monitoring is one of them. In order to achieve the complete non-intrusive concept and to adapt to the changes in the environment, this paper proposes the adaptive non-intrusive load monitoring system framework that applied in the monitoring system, taking low frequency acquisition and steady-state feature extraction for reducing its setup costs. The method adopts unsupervised learning, which builds classifier in load state by Gaussian mixture model (GMM)/ Sequential Expectation-maximization (SEM) and does adaptive fine-tuning for the system by online data. The results show that the framework can adapt the changes in the environment and detect new unknown state for providing a more complete on-line monitoring system solution.

Smart grid is one of the most important applications of the Internet of Things (IoT) for environmental sustainability and energy efficiency issues in recent years. IP-based wireless sensor networks (IP-WSNs) are considered as one of the promising wireless communication technologies applied in smart grid for providing pervasive communications and control capabilities at low cost as well as connecting metering devices with wired area network (WAN) infrastructures seamlessly without a requirement for deploying proxies. In addition, the IPv6 Internet has become an inevitable trend for all-IP communication because of its large address space and others advantages over IPv4. One of the main challenges for connecting WSNs and IPv6 Internet is IPv6 address configuration since nodes with unique address are a prerequisite for reliability and end to end communication. Hence, an IPv6 address configuration scheme called MPIPA is proposed in this paper. MPIPA utilizes three-dimensional locations coordinates to assign each node a unique spatial IPv6 address based on grouping methods and scan-line scheme. Besides, Assignment Success Rate (ASR) is used in this paper to evaluate the probability that assigns unique IP address to nodes successfully. The simulation results show that MPIPA enables over 8, 000 nodes to be assigned IP address successfully when ASR is still nearly 90%.

SVG (scalable vector graphics) has become the standard format for 2D graphics in HTML5. Although some image-to-SVG conversion systems had been proposed, the sizes of files they produced are still large. In [1], we proposed a new system to convert raster comic images into vector SVG files. The compression ratio is better than the previous methods. However, these methods do not process text in raster images. In this paper, we improve our system to recognize text elements in the comic and use these text elements to provide better compression and novel applications. The proposed method uses SCW (sliding concentric windows) and SVM (support vector machine) to identify text regions. Then, OCR (optical character recognition) is applied to recognize text elements in those regions. Instead of encoding the text regions as vectors, the text elements are embedded in the SVG file along with their coordinate values. Experimental results show that we can reduce the file sizes to about 52% of the original SVG files. Using these text elements, we can translate comics into other languages to provide multilingual services easily. Text/content-based image search can be supported efficiently. It can also provide a novel application system for story teller.

Wireless Sensor Networks (WSNs) consist of group sensor nodes which are placed in an area to monitor the changes of environment. Usually, sensing data are gathered and stored in a data server which maintains a database to organize and manage numerous of WSNs data. It allows researchers to retrieve these data for further study or analysis. Since the size of WSNs data is huge and the storage resource is limited, this database needs compression to lower the data size. In this paper, we propose a video-like compression method for high efficiency database storage of WSNs. First, the raw data are arranged according to the spatial correlation as an image frame. Then, several image frames with temporal correlation are maintained as a sequence of frames and lossless video compression is adopt for lowering the data size. Based on this idea, we also propose a data retrieve/query algorithm for parallel processing. The trade-off between space saving and query time is discussed after experiencing with real-world data. At last, we compare our proposed method to MySQL, a well-known database which compression is supported. The experimental results reveal that our method achieves over 96% of the space savings. It is over 13% more than that achieved by MySQL.

Intelligent home appliances and friendly human-machine interface design can provide a more comfortable living space for residents in smart home. However, excessive home automation sometimes may lose the sense of reality on operating home appliances. Virtual objects in smart home are not real home furnishings, and equipments, but it has the ability of actual home appliances or equipment to provide relevant home services through ubiquitous computing, virtual human-machine interface. That is, virtual objects retain the original features of appliance operation without losing high-technical and home-automation utilization on it. In this paper, we design a virtual object being friendly human-machine interaction to control the music playback service in smart home. The virtual object is called virtual compact disc (CD) album, virtual-CD. The residents can take a virtual-CD, and do some operating gestures and entity movements on it, and then the music will be played, paused, and stopped according to various operations of virtual object in smart home. By the way, the resident can easily enjoy high-technical home service without losing traditional appliance operation in smart home.

Traffic smoothing is an efficient means to reduce the bandwidth requirement for transmitting a variable-bit-rate video stream. Several traffic-smoothing algorithms have been presented to offline compute the transmission schedule for a prerecorded video. For live video applications, Sen present a sliding-window algorithm, referred to as SLWIN(k), to online compute the transmission schedule on the fly. SLWIN(k) looks ahead W video frames to compute the transmission schedule for the next k frametimes, where klesw. Note that W is upper bounded by the initial delay of the transmission. The time complexity of SLWIN(k) is O(W*N/k) for an N frame live video. In this paper, we present an O(N) online traffic-smoothing algorithm and two variants, denoted as FOS, FOS1 and FOS2, respectively. Note that O(N) is a trivial lower bound of the time complexity of the traffic-smoothing problem. Thus, the proposed algorithm is optimal. We compare the performance of our algorithms with SLWIN(k) based on several benchmark video clips. Experiment results show that FOS2, which adopts the aggressive workahead heuristic, further reduces the bandwidth requirement and better utilizes the client buffer for real-time interactive applications in which the initial delays are small

Active queue management (AQM) schemes have motivated many researchers to investigate more effective methods to control network congestion. Most AQM schemes are evaluated by their designers on the basis of router-centric metrics, such as queuing delay, link utilization and packet drop ratio. These metrics are important to network operators but they may not reflect the quality of service delivered to end-users. In this paper we propose a method aimed to provide users with better services in terms of end-to-end delay and packet loss ratio. The method captures a significant traffic increase at an early stage and signals TCP sources to slow down. In this way, TCP can quickly adjust the transmission rate, and thereby prevent overloading the network. In addition to RED, another two prominent AQM schemes, namely, random early marking (REM) and adaptive RED (ARED), are compared with the proposed method. Simulation results show that under various network loads and a range of network propagation delays, our method can achieve lower end-to-end delay and packet loss ratio, compared with all aforementioned schemes.

Previous real-time disk scheduling algorithms assume that each disk request incurs a physical disk mechanical operation and only consider how to move the disk head under real-time constraints. However, with the increased capacity of on-disk cache, modern disk drives read-ahead data aggressively. Thus, the on-disk cache may service many disk requests without incurring physical disk access. By exploring the design methodology of on-disk cache, in this paper, we propose cache-aware real-time disk scheduling algorithms that take the on-disk cache into consideration during scheduling. Therefore, the scheduling algorithm can help the cache replacement scheme to minimize the cache miss ratio. Besides, the service timing estimation is more accurate in schedulability analysis since the cache effect is considered during scheduling. A simulation-based evaluation shows the proposed scheduling algorithms to be highly successful as compared with the classical real-time disk scheduling algorithms. For example, under sequential workload with 10 sequential streams, the data throughput of our scheme is 1.1 times that of DM-SCAN.

In this paper, we study the problem of designing an efficient ARQ algorithm for supporting real-time video streaming applications. This problem differs from traditional non-real-time error control problem in which a late arrival packet can seriously degrade the video playback quality. It is interesting to notice that if a lost packet is detected while the client buffer is running at a high position, then the probability to recover the lost packet is also high. In this paper, we present a new error control algorithm, called BREC (buffer-controlled retransmission-based error control), to dynamically control the buffer running at a specific level. By the experiments on a true VOD system in which we implement the BREC scheme, the packet loss rate can be reduced by an order of magnitude. We also show that the performance can be further improved by detecting the rate of changes in running buffer positions and use it to enhance the ability of buffer position control. Both analytical model and experimental results show that our mechanism significantly decrease packet loss and improves QoS even if a client, such as set-top box, PDA (personal data assistant) or cellular phone, preserves only a limited amount of playback buffer (200K bytes of memory is used in our experiments).

Due to advances in network technologies, providing streaming services over the Internet has gained in popularity. Because a video stream is in compressed format, it is naturally with the variable bit rate (VBR) property and its traffic is highly burst. With the installation of a video proxy between access networks (e.g. local area networks, LAN) and backbone networks (e.g. wide area networks, WAN), video staging is proposed to cache part of the requested video into the video proxy close to clients. In this mechanism, the video can be streamed using a constant bit rate (CBR) network service across the backbone WAN and the WAN bandwidth requirement can be significantly reduced. In this paper, we propose a new approach, called "caching selected after smoothing" (CSAS), to handling video staging. The aim of the CSAS algorithm is to integrate our previous video caching and smoothing technique so as to reduce the required WAN bandwidth even more. Experiments on benchmark videos using different evaluation indices, including the proxy storage requirement, the WAN bandwidth requirement, and the WAN bandwidth utilization, show that our approach is more effective than the conventional "cut after smoothing" (CAS) algorithm.

In a multimedia system, disk I/O subsystem is the most important component due to its relatively limited throughput and large delay. Previously, by applying SCAN to reschedule tasks having the same deadline, SCAN-EDF tries to improve disk throughput while real time constraints can be satisfied. In DM-SCAN, groups of tasks that can be successfully rescheduled by SCAN under specified real time requirements are identified. They are called MSGs (maximum-scannable-groups). An enlarged-MSG (E-MSG) is proposed to further expand the MSG concept and thus to obtain more improvement in disk throughput. By removing some excess constraints on MSG, E-MSG merges several MSGs as a new scannable group. Experimental results show that the E-MSG scheme is better than both SCAN-EDF and MSG in the disk throughput obtained.

Traffic-smoothing for delivery of online VBR media is one of the most important problems when designing streaming multimedia applications. Given the available client buffer b and playback delay D, Rexford et al. (1997) introduced a window-based approach called SLWIN(k) to smooth online generated traffic for the pre-specified window size W (W

Real-time disk scheduling is one of the most important problems in designing a multimedia system. It has been proved to be NP-complete. Recently, various approaches have been proposed to improve disk throughput under guaranteed real-time requirements. SCAN-EDF, which scans the disk surface to retrieve the task data block under the disk head in order to re-schedule tasks in a real-time EDF (earliest deadline first) schedule, is one of the best-known real-time disk scheduling methods. Since tasks rescheduled in SCAN-EDF should have the same deadline, its efficiency depends on the number of tasks with the same deadline. If all tasks have different deadlines, the scheduling results of SCAN-EDF would be the same as EDF. In this paper, we improve SCAN-EDF by applying different hybrid local-merging and global-inserting schemes. As opposed to SCAN-EDF, in our method tasks rescheduled by SCAN may have different deadlines. Its efficiency is not limited by the number of tasks that have the same deadlines. Experiments show that the proposed method is significantly better than SCAN-EDF. In terms of disk throughput, the improvement obtained is 24% greater than the best-known SCAN-EDF method.

Traffic smoothing for delivery of online VBR media streams is one of the most important problems in designing multimedia systems. Given available client buffer and a window-sliding size, conventional approaches try to reduce bandwidth allocated in each window. However, they can not lead to the minimization of bandwidth allocated for transmitting the entire stream. Although a window-sliding approach was introduced previously to further reduce the bandwidth allocated, it was computational costly. In this paper, an effective and efficient online traffic-smoothing scheme is proposed. Different from the conventional static window-sliding approaches, this approach dynamically decides the suitable window-sliding size to online smooth the bursty traffic. Then, an aggressive workahead scheme is applied in transmitting the entire stream. By examining different media streams, the approach has a small bandwidth, high bandwidth utilization and small computation cost. Considering the online transmission of a Star War movie, our approach result is 13% less for the bandwidth and 4% less for the network idle rate than SLWIN(1). Comparing the number of window sliding, our approach is 75% less than SLWIN(1). The relations between the characteristic of the input traffic and the behavior of obtained scheduling results are discussed. Finally, an extension of the proposed approach to resolve the latency and quality tolerance applications is also introduced.

Given a pre-recorded VBR video, we have proposed an O(n) algorithm to smooth the transmission schedule with the minimum required resources. N is the number of video frames. As n is usually very large and varying for different videos, it is not suitable for online computation. To facilitate resource management and admission control for QoS (quality-of-service) guarantees, we need to explore the relations among the required resources. Thus, whenever a new request is presented, the admission control procedure can easily check the required resources against the available resources and decides to admit this new request or not. To compute these relations (such as rate-buffer and rate-delay), a native algorithm takes O(n/sup 3/) time complexity. An O(n log n) algorithm is proposed to characterize the low-bounds of resources allocated for transmitting a pre-recorded VBR video. Having these pre-computed functions, the admission control procedure is as simple as a chart look-up with O(1) time complexity to allocate the required resources.

Real-time disk scheduling is important to multimedia systems support for digital audio and video. In these years, various approaches are presented to use the seek-optimizing scheme to improve the disk throughput of a real-time guaranteed schedule. However, as these conventional approaches apply SCAN only to the requests with the same deadline or within the same constant-sized group, their improvements are limited. In this paper, we introduce the DM-SCAN (deadline-modification-SCAN) algorithm with an idea of MSG (maximum-scannable-group). The proposed DM-SCAN method can apply SCAN to MSG iteratively by modifying request deadlines. We have implemented the DM-SCAN algorithm on UnixWare 2.01. Experiments show that DM-SCAN is significantly better than that of the best-known SCAN-EDF method in both the obtained disk throughput and the number of supported disk requests.

Query-based learning (QBL) has been introduced for training a supervised network model with additional queried samples. Experiments demonstrated that the classification accuracy is further increased. Although QBL has been successfully applied to supervised neural networks, it is not suitable for unsupervised learning models without external supervisors. In this paper, an unsupervised QBL (UQBL) algorithm using selective-attention and self-regulation is proposed. Applying the selective-attention, we can ask the network to respond to its goal-directed behavior with self-focus. Since there is no supervisor to verify the self-focus, a compromise is then made to environment-focus with self-regulation. In this paper, we introduce UQBL1 and UQBL2 as two versions of UQBL; both of them can provide fast convergence. Our experiments indicate that the proposed methods are more insensitive to network initialization. They have better generalization performance and can be a significant reduction in their training size.

In this paper, a three-layer force-directed self-organizing map is designed to resolve the circuit placement problem with arbitrarily shaped rectilinear modules. The proposed neural model with an additional hidden layer can easily model a rectilinear module by a set of hidden neurons to correspond the partitioned rectangles. With the collective computing from hidden neurons, these rectilinear modules can correctly interact with each other and finally converge to a good placement result. In this paper, multiple contradictory criteria are accounted simultaneously during the placement process, in which, both the wire length and the module overlap are reduced. The proposed model has been successfully exploited to solve the time consuming rectilinear module placement problem. The placement results of real rectilinear test examples are presented, which demonstrate that the proposed method is better than the simulated annealing approach in the total wire length. The appropriate parameter values which yield good solutions are also investigated.

In this paper, a three-layer self-organizing neural network is designed to resolve the cell placement problem with arbitrarily-shaped rectilinear cells. The proposed model with additional hidden layer can easily model the rectilinear cells by a set of hidden neurons which correspond to the partitioned rectangles of cells, called "molecule model". With the collective computing property and solid-connection of neural networks, these rectilinear cells can correctly interact with each other and finally converge to a good placement result. The placement results of rectilinear test examples have been presented In this paper, multiple contradictory criteria are accounted simultaneously during the placement process, in which, total wire length is reduced without module overlap.>